Feature tokenizer-graph attention network (ft-GAT) model-based electronic device for predicting spontaneous breathing success rate of artificial respiratory patient, and control method therefor

Abstract

Disclosed are an electronic device and a control method therefor. The electronic device according to the present disclosure includes: a memory storing a feature tokenizer-graph attention network (FT-GAT) model for obtaining a prediction value of a spontaneous breathing success rate of a patient on the basis of patient information; and a processor connected to the memory and identifying the prediction value of the spontaneous breathing success rate of the patient by inputting at least one feature value of the patient to the FT-GAT model, wherein the FT-GAT model may include: a tokenizer module that converts a feature value corresponding to the patient information into an embedding value in a latent space; a multi-head graph attention module that updates a node corresponding to the feature value; and a multi-layer perceptron (MLP) module that outputs the prediction value of the spontaneous breathing success rate of the patient, corresponding to the embedding value, by applying the updated node.

Description

Electronic device for predicting the success rate of spontaneous breathing in a ventilated patient based on the FT-GAT (FEATURE TOKENIZER - GRAPH ATTENTION NETWORK) model and its control method

[0001] The present disclosure relates to an electronic device for predicting the success rate of spontaneous breathing of a patient and a control method thereof, and more particularly, to an electronic device for predicting the success rate of spontaneous breathing of a patient receiving artificial respiration using patient information based on an FT-GAT (Feature Tokenizer - Graph Attention Network) model and a control method thereof.

[0002] Mechanical ventilation (MV) is a treatment method that temporarily replaces mechanical ventilation to improve the symptoms of patients with acute respiratory failure. Previous statistics show that 39% of patients admitted to intensive care units (ICUs) require invasive ventilation. The COVID-19 pandemic has highlighted the importance of oxygen therapy and respiratory support machines, including intravascular mechanical ventilation, in patients with acute respiratory distress syndrome.

[0003] However, mechanical ventilation poses additional risks to patients, as it can lead to long-term complications, including ventilator-associated pneumonia, diaphragmatic and muscle weakness, laryngeal damage, and dysphagia. Furthermore, long-term mechanical ventilation is associated with increased mortality, prolonged dependence on the healthcare system after survival, psychological problems, and financial burden. To minimize complications, physicians should attempt to wean patients from mechanical ventilation as soon as they are ready. This approach will also help reallocate medical resources to other critically ill patients, given the limited capacity of intensive care units.

[0004] To wean a patient from mechanical ventilation, physicians typically follow a three-step approach. The first step is to assess whether the patient meets the criteria for "readiness for mechanical ventilation" based on vital signs, mental status, improvement in condition, oxygen requirements, and ventilator settings. The physician then performs a spontaneous breathing trial (SBT), which assesses the patient's ability to breathe with minimal or no ventilator support for 30 minutes to 2 hours. SBT methods include a T-piece trial without ventilator support and low-volume ventilator support. Low-volume ventilator support for SBT typically includes pressure-support ventilation (PSV) of up to 5-7 cmH2O, continuous positive airway pressure (CPAP) alone (i.e., no inspiratory pressure support), with or without positive end-expiratory pressure (PEEP), or automatic tube compensation. The success of spontaneous breathing attempts provides objective insight into the feasibility of endotracheal intubation, which is the final step toward weaning off mechanical ventilation.

[0005] Although many researchers have focused on finding robust predictors of successful mechanical ventilation weaning, such as the Rapid Shallow Breathing Index (RSBI), Heart Rate Variability (HRV), or peak inspiratory and expiratory pressures, no single robust predictor has yet been identified. In practice, the decision to initiate mechanical ventilation weaning is based on the multifaceted judgment of the intensive care physician, based on various clinical circumstances and the ventilation weaning protocol. However, 25-40% of patients fail the first spontaneous ventilation attempt, and 10-25% experience extubation failure, with this rate varying across institutions and physicians.

[0006] Artificial intelligence (AI) is being developed for various medical support systems, including intensive care units (ICUs). In Korea, despite the efforts of many ICU specialists, weaning critically ill patients from mechanical ventilation relies heavily on physicians, not ICU specialists or respiratory care specialists. This often leads to early discharge or prolonged hospitalization. Therefore, we aim to develop a more useful deep learning model that supports clinician decision-making and enables early and appropriate removal from mechanical ventilation.

[0007] This study is a two-stage process, the first of which aims to predict the success of spontaneous ventilation attempts. Furthermore, the study was systematically conducted using the Joanna Briggs Institute (JBI) Critical Assessment Checklist (Appendix Table 1). This is the first study to convert and analyze feature values ​​(which can be factors or variables) contained in tabular data into graphs reflecting the perspectives of pulmonologists.

[0008] The purposes of the present disclosure are not limited to those mentioned above, and other purposes and advantages of the present disclosure not mentioned above can be understood through the following description and will be more clearly understood through the embodiments of the present disclosure. Furthermore, it will be readily apparent that the purposes and advantages of the present disclosure can be realized by the means and combinations thereof set forth in the claims.

[0009] Artificial intelligence (AI) is being developed for various medical support systems, including intensive care units (ICUs). In Korea, despite the efforts of many ICU specialists, weaning critically ill patients from mechanical ventilation relies heavily on physicians, not ICU specialists or respiratory care specialists. This often leads to early discharge or prolonged hospitalization. Therefore, we aim to develop a more useful deep learning model that supports clinician decision-making and enables early and appropriate removal from mechanical ventilation.

[0010] This study is a two-stage process, the first of which aims to predict the success of spontaneous ventilation attempts. Furthermore, the study was systematically conducted using the Joanna Briggs Institute (JBI) Critical Assessment Checklist (Appendix Table 1). This is the first study to convert and analyze feature values ​​(which can be factors or variables) contained in tabular data into graphs reflecting the perspectives of pulmonologists.

[0011] The purposes of the present disclosure are not limited to those mentioned above, and other purposes and advantages of the present disclosure not mentioned above can be understood through the following description and will be more clearly understood through the embodiments of the present disclosure. Furthermore, it will be readily apparent that the purposes and advantages of the present disclosure can be realized by the means and combinations thereof set forth in the claims.

[0012] Based on patient information, we can develop a useful deep learning model to predict the probability of spontaneous breathing, which can lead to appropriate and early release from mechanical ventilation, to assist clinicians in decision-making.

[0013] Aspects, features and advantages of specific embodiments of the present disclosure will become more apparent from the following description taken in conjunction with the accompanying drawings.

[0014] FIG. 1 is a block diagram illustrating the configuration of an electronic device according to an embodiment of the present disclosure.

[0015] FIG. 2 is a block diagram illustrating an FT-GAT model including a tokenizer module, a multi-head graph attention module, and a multi-layer perceptron module according to an embodiment of the present disclosure.

[0016] FIG. 3 is a table showing patient information and general information related to mechanical ventilation according to one embodiment of the present disclosure.

[0017] FIG. 4 is a flowchart illustrating patient information and success and failure related to spontaneous breathing attempts in an intensive care unit according to one embodiment of the present disclosure.

[0018] FIG. 5 is a diagram illustrating the process of failure and success of an attempt at self-breathing according to one embodiment of the present disclosure.

[0019] FIG. 6 is a diagram illustrating the operation of a tokenizer that converts patient information into an embedding value according to one embodiment of the present disclosure.

[0020] FIG. 7 is a diagram illustrating a plurality of nodes corresponding to feature values ​​of patient information and edges connecting the plurality of nodes, according to one embodiment of the present disclosure.

[0021] FIG. 8 is a diagram illustrating an attention mechanism for identifying an attention coefficient for each node corresponding to a feature value of patient information according to one embodiment of the present disclosure.

[0022] FIG. 9 is a diagram for explaining an operation of updating each node corresponding to a feature value of patient information based on an adjacent node and an attention coefficient of the adjacent node, according to one embodiment of the present disclosure.

[0023] FIG. 10 is a table illustrating hyperparameters required to predict the success rate of an autonomous breathing attempt using the FT-GAT model according to one embodiment of the present disclosure.

[0024] FIG. 11 is a table showing patient information related to predicting the success rate of spontaneous breathing attempts using the FT-GAT model according to one embodiment of the present disclosure.

[0025] FIG. 12 is a diagram showing experimental data evaluating the performance of the FT-GAT model in predicting the success rate of autonomous breathing attempts according to one embodiment of the present disclosure.

[0026] FIG. 13 is a table showing experimental data evaluating the performance of the FT-GAT model in predicting the success rate of autonomous breathing attempts according to one embodiment of the present disclosure.

[0027] FIG. 14 is a table showing indicators for explaining performance comparison between the FT-GAT model and other models according to one embodiment of the present disclosure.

[0028] FIG. 15 is a table showing the experimental results for each structure of each node included in the FT-GAT model according to one embodiment of the present disclosure.

[0029] FIG. 16 is a diagram showing a comparison of ROC curves of the FT-GAT model and other conventional models according to one embodiment of the present disclosure.

[0030] FIG. 17 is a diagram showing an average heatmap of normalized attention coefficients according to one embodiment of the present disclosure.

[0031] FIG. 18 is a table showing data representing the performance of an FT-GAT model according to one embodiment of the present disclosure.

[0032] The present embodiments may be modified and have various embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the scope to specific embodiments, but should be understood to encompass various modifications, equivalents, and / or alternatives of the embodiments of the present disclosure. In connection with the description of the drawings, similar reference numerals may be used for similar components.

[0033] In describing the present disclosure, if it is determined that a specific description of a related known function or configuration may unnecessarily obscure the gist of the present disclosure, a detailed description thereof will be omitted.

[0034] Additionally, the following embodiments may be modified in various other forms, and the scope of the technical concepts of the present disclosure is not limited to the following embodiments. Rather, these embodiments are provided to further faithfully and completely convey the technical concepts of the present disclosure to those skilled in the art.

[0035] The terminology used in this disclosure is for the purpose of describing specific embodiments only and is not intended to limit the scope of the rights. Singular expressions include plural expressions unless the context clearly dictates otherwise.

[0036] In this disclosure, expressions such as “has,” “can have,” “includes,” or “may include” indicate the presence of a corresponding feature (e.g., a component such as a number, function, operation, or part), and do not exclude the presence of additional features.

[0037] In this disclosure, expressions such as “A or B,” “at least one of A and / or B,” or “one or more of A or / and B” can include all possible combinations of the listed items. For example, “A or B,” “at least one of A and B,” or “at least one of A or B” can all refer to (1) including at least one A, (2) including at least one B, or (3) including both at least one A and at least one B.

[0038] The expressions “first,” “second,” “first,” or “second,” etc., used in this disclosure can describe various components, regardless of order and / or importance, and are only used to distinguish one component from another, but do not limit the components.

[0039] When it is said that a component (e.g., a first component) is “(operatively or communicatively) coupled with / to” or “connected to” another component (e.g., a second component), it should be understood that the component may be directly coupled to the other component, or may be connected through another component (e.g., a third component).

[0040] On the other hand, when it is said that a component (e.g., a first component) is "directly connected" or "directly connected" to another component (e.g., a second component), it can be understood that no other component (e.g., a third component) exists between the component and the other component.

[0041] The expression "configured to" as used in the present disclosure may be used interchangeably with, for example, "suitable for," "having the capacity to," "designed to," "adapted to," "made to," or "capable of." The term "configured to" may not necessarily mean only "specifically designed to" in terms of hardware.

[0042] Instead, in some contexts, the phrase "a device configured to" may mean that the device, in conjunction with other devices or components, is "capable of" performing A, B, and C. For example, the phrase "a processor configured (or set) to perform A, B, and C" may refer to a dedicated processor (e.g., an embedded processor) for performing those operations, or a general-purpose processor (e.g., a CPU or application processor) that can perform those operations by executing one or more software programs stored in a memory device.

[0043] In the embodiments, a 'module' or 'part' performs at least one function or operation, and may be implemented as hardware or software, or as a combination of hardware and software. Furthermore, a plurality of 'modules' or 'parts' may be integrated into at least one module and implemented as at least one processor, except for a 'module' or 'part' that needs to be implemented as a specific hardware.

[0044] Meanwhile, the various elements and areas in the drawings are schematically drawn. Therefore, the technical concept of the present invention is not limited by the relative sizes or spacing depicted in the attached drawings.

[0045] Hereinafter, with reference to the attached drawings, embodiments according to the present disclosure will be described in detail so that a person having ordinary knowledge in the technical field to which the present disclosure pertains can easily implement the present disclosure.

[0046] The electronic device (100) may be a device that trains an artificial intelligence model, a machine learning model, a deep learning model, etc. for predicting a patient's self-breathing success rate based on learning data, and inputs patient information, characteristic information of the patient information, etc. into the trained artificial intelligence model, machine learning model, or deep learning model to obtain a prediction value of the patient's self-breathing success rate.

[0047] The electronic device (100) may be, for example, a server, a computer that provides services to clients via a network. The server may be an FTP server, a web server, a database server, or a cloud-based server, and may be built with an operating system such as Linux.

[0048] In addition, the electronic device (100) may be a computing device capable of performing data operation, such as a smartphone, a tablet personal computer, a laptop personal computer, a netbook computer, a mobile device, a wearable device, etc.

[0049] An electronic device (100) according to an embodiment of the present disclosure is not limited to the above-described devices, and the electronic device (100) may be implemented as an electronic device (100) having two or more functions of the above-described devices.

[0050] FIG. 1 is a block diagram illustrating the configuration of an electronic device (100) according to one embodiment of the present disclosure.

[0051] Referring to FIG. 1, the electronic device (100) may include a memory (110) and a processor (120). However, the configuration of the electronic device (100) is not limited thereto, and may additionally include a communication interface, a user interface, a display, etc., or may omit some configurations.

[0052] The memory (110) temporarily or non-temporarily stores various programs or data, and transmits the stored information to the processor (120) upon a call from the processor (120). In addition, the memory (110) can store various information necessary for calculations, processing, or control operations of the processor (120) in an electronic format.

[0053] The memory (110) may include, for example, at least one of a main memory and an auxiliary memory. The main memory may be implemented using a semiconductor storage medium such as ROM and / or RAM. The ROM may include, for example, a conventional ROM, EPROM, EEPROM, and / or MASK-ROM. The RAM may include, for example, DRAM and / or SRAM. The auxiliary memory may be implemented using at least one storage medium capable of permanently or semi-permanently storing data, such as a flash memory device, an SD (Secure Digital) card, a solid state drive (SSD), a hard disk drive (HDD), an optical media such as a magnetic drum, a compact disc (CD), a DVD, or a laser disc, a magnetic tape, a magneto-optical disc, and / or a floppy disk.

[0054] The memory (110) can store artificial intelligence models, machine learning models, and deep learning models. Specifically, the memory (110) can store nodes, layers, edges, weights, biases, activation functions, loss functions, and backpropagation coefficients. Additionally, the memory (110) can store training data required for model training.

[0055] The memory (110) can store an FT-GAT model for obtaining a patient's spontaneous breathing success rate based on patient information.

[0056] The memory (110) can store information about a tokenizer (210) module that converts feature values ​​corresponding to patient information into embedding values ​​in a latent space, a multi-head graph attention (220) module that updates nodes corresponding to feature values, and a multi-layer perceptron (230) (MLP, Multi Layer Perceptron) module that outputs a prediction value of a patient's spontaneous breathing success rate corresponding to the embedding value by applying the updated nodes.

[0057] The memory (110) can store information about the tokenizer (210) module, the multi-head graph attention (220) module, and the multi-layer perceptron (230) module corresponding to the FT-GAT model, for example, various nodes, edges, layers, coefficients, parameters, functions, etc. required for the control operation of each module.

[0058] The memory (110) can store information about the structure of each module, graph structure, etc., and can store instructions required for controlling and operating each module.

[0059] The memory (110) can store patient information, characteristic values ​​of the patient information, and a prediction value of the success rate of spontaneous breathing obtained based on the patient information.

[0060] The processor (120) controls the overall operation of the electronic device (100). Specifically, the processor (120) is connected to the configuration of the electronic device (100) including the memory (110) as described above, and can control the overall operation of the electronic device (100) by executing at least one instruction stored in the memory (110) as described above. In particular, the processor (120) may be implemented as one processor or may be implemented as multiple processors.

[0061] The processor (120) may be implemented in various ways. For example, one or more processors (120) may include one or more of a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), an Accelerated Processing Unit (APU), a Many Integrated Core (MIC), a Digital Signal Processor (DSP), a Neural Processing Unit (NPU), a hardware accelerator, or a machine learning accelerator. The one or more processors (120) may control one or any combination of other components of the electronic device (100) and perform operations related to communication or data processing. The one or more processors (120) may execute one or more programs or instructions stored in the memory (110). For example, the one or more processors (120) may perform a method according to an embodiment of the present disclosure by executing one or more instructions stored in the memory (110).

[0062] When a method according to an embodiment of the present disclosure includes a plurality of operations, the plurality of operations may be performed by one processor (120) or may be performed by a plurality of processors (120). For example, when a first operation, a second operation, and a third operation are performed by a method according to an embodiment, the first operation, the second operation, and the third operation may all be performed by the first processor, or the first operation and the second operation may be performed by the first processor (e.g., a general-purpose processor) and the third operation may be performed by the second processor (e.g., an artificial intelligence-only processor).

[0063] One or more processors (120) may be implemented as a single core processor including one core, or may be implemented as one or more multicore processors including multiple cores (e.g., homogeneous multicores or heterogeneous multicores). When one or more processors (120) are implemented as multicore processors, each of the multiple cores included in the multicore processor may include an internal memory of the processor, such as an on-chip memory (110), and a common cache shared by the multiple cores may be included in the multicore processor (120). In addition, each of the multiple cores (or some of the multiple cores) included in the multicore processor (120) may independently read and execute a program instruction for implementing a method according to an embodiment of the present disclosure, or all (or some) of the multiple cores may be linked to read and execute a program instruction for implementing a method according to an embodiment of the present disclosure.

[0064] When a method according to an embodiment of the present disclosure includes a plurality of operations, the plurality of operations may be performed by one core among the plurality of cores included in a multi-core processor, or may be performed by the plurality of cores. For example, when a first operation, a second operation, and a third operation are performed by a method according to an embodiment, the first operation, the second operation, and the third operation may all be performed by a first core included in the multi-core processor, or the first operation and the second operation may be performed by a first core included in the multi-core processor, and the third operation may be performed by a second core included in the multi-core processor.

[0065] In embodiments of the present disclosure, the processor (120) may mean a system on a chip (SoC) in which one or more processors (120) and other electronic components are integrated, a single core processor, a multi-core processor, or a core included in a single core processor or a multi-core processor, wherein the core may be implemented as a CPU, a GPU, an APU, a MIC, a DSP, an NPU, a hardware accelerator, or a machine learning accelerator, but embodiments of the present disclosure are not limited thereto.

[0066] The processor (120) is connected to the memory (110) and can input at least one feature value of the patient into the FT-GAT model to identify a prediction value of the patient's spontaneous breathing success rate.

[0067] FIG. 2 is a block diagram illustrating an FT-GAT model including a tokenizer (210) module, a multi-head graph attention (220) module, and a multi-layer perceptron (230) module according to one embodiment of the present disclosure.

[0068] Referring to FIG. 2, the processor (120) can input feature values ​​corresponding to patient information into the tokenizer (210) module to identify embedding values ​​in the latent space.

[0069] The processor (120) can control the multi-head graph attention (220) module to update a node corresponding to a feature value.

[0070] The processor (120) can apply the updated node to the multi-layer perceptron (230) (MLP, Multi Layer Perceptron) module.

[0071] The processor (120) can obtain a prediction value for the patient's spontaneous breathing success rate by inputting the embedding value into a multi-layer perceptron (230) module to which the updated node is applied.

[0072] A more specific description of the electronic device (100) control and FT-GAT model control operation of the processor (120) is described with reference to FIGS. 3 to 18.

[0073] FIG. 3 is a table showing patient information and general information related to mechanical ventilation according to one embodiment of the present disclosure.

[0074] Referring to Figure 3, the electronic medical records of registered patients were reviewed, and the collected input characteristics are as shown in Figure 3.

[0075] Specifically, all input features obtained immediately before spontaneous ventilation attempt included age, sex, body mass index (BMI), mechanical ventilation mode, and mechanical ventilation parameters (pressure support (PS) level, vital capacity (FiO2), and tidal volume). In addition, the ratio of arterial oxygen pressure (PaO2) to FiO2 (P / F), pCO2, tidal volume - 6 mL / kg predicted body weight (PBW), mechanical ventilation duration, heart rate, respiratory rate, and Glasgow Coma Score (GCS). Mechanical ventilation modes consisted of assist control (AC), synchronized intermittent mandatory ventilation (SIMV), PSV, and CPAP. In addition, the TV-6 mL / kg PBW was calculated using the average of the three tidal volume values.

[0076] FIG. 4 is a flowchart illustrating patient information and success and failure related to spontaneous breathing attempts in an intensive care unit according to one embodiment of the present disclosure.

[0077] Referring to Figure 4, one can see the organization and flow of data related to patient information and success and failure of spontaneous breathing attempts in the intensive care unit.

[0078] FIG. 5 is a diagram illustrating the process of failure and success of an attempt at self-breathing according to one embodiment of the present disclosure.

[0079] Referring to Figure 5, in the additional collection procedure for cases of successful or unsuccessful spontaneous breathing, the dotted lines indicate additional collection points if (a) the time between intubation and the failed spontaneous breathing attempt exceeded 48 hours or (b) the failed spontaneous breathing attempt exceeded 48 hours.

[0080] The dataset was preprocessed for data analysis. The gender feature was encoded using one-hot encoding, and the mechanical ventilation mode feature was encoded using label encoding, from weakest to strongest respiratory support, as follows: CPAP (0), PSV (1), SIMV (2), AC (3). Additionally, min-max scaling was performed on unencoded numerical features to obtain values ​​between 0 and 1. The mathematical formula for min-max scaling is shown in Equation 1 below.

[0081]

[0082] Here, max(X) and min(X) represent the maximum and minimum values ​​of feature X, respectively.

[0083] To design the FT-GAT model, we designed all features into a graph structure, inspired by graph neural networks (GNNs), which aggregate feature information from adjacent nodes. The graphs used in the analysis were structured by two respiratory medicine specialists with over 6 and 15 years of experience, respectively. Furthermore, we designed a novel spontaneous breathing success rate prediction model called FT-GAT, which performs graph analysis by utilizing a graph network (GAT) structure that includes an attention mechanism.

[0084] The processor (120) is a feature value x (categorical x) corresponding to patient information. cat and numeric x num ) is controlled by the tokenizer (210) module included in the FT-GAT model to convert it into an embedding value.

[0085] FIG. 6 is a diagram for explaining the operation of a tokenizer (210) that converts patient information into an embedding value according to one embodiment of the present disclosure.

[0086] Referring to Figure 6, given feature x jThe embedding value for is calculated using the following mathematical formula 2.

[0087]

[0088]

[0089] W j is the weight corresponding to the jth feature value, and B j is the bias corresponding to the jth feature value. e T j represents a one-hot vector corresponding to the j-th category feature.

[0090] The processor (120) can obtain an embedding value by multiplying a weight and adding a bias to a feature value corresponding to patient information. Here, the feature value may be at least one of a category corresponding to patient information and a quantitative value corresponding to patient information.

[0091] In various embodiments, when embedding values ​​in the latent space corresponding to the feature values ​​of patient information are identified, the processor (120) can identify the position of each embedding value in the latent space and the distance between them. The processor (120) can identify the embedding value with the largest average value of distances from other embedding values ​​in the latent space among the embedding values ​​corresponding to a plurality of different feature values ​​as the noise embedding value. The processor (120) can delete the noise embedding value and delete the node corresponding to the noise embedding value. Accordingly, the processor (120) can perform more accurate model calculations by excluding values ​​that can be considered close to noise due to low correlation with other feature values ​​in the patient information.

[0092] FIG. 7 is a diagram illustrating a plurality of nodes corresponding to feature values ​​of patient information and edges connecting the plurality of nodes, according to one embodiment of the present disclosure.

[0093] Referring to Fig. 7, a plurality of nodes corresponding to feature values ​​of patient information and a plurality of nodes can be connected to each other by having a specific structure.

[0094] The processor (120) can input the characteristic values ​​of patient information into each node as above to obtain output values.

[0095] The configuration of the multi-head graph attention (220) module of FT-GAT includes multiple single-graph attention layers. The single-graph attention layer is a node set H = {H1, H2,..., H N}, H i ∈ R d It takes as input N, where N represents the number of nodes (i.e., features), and d represents the dimension of each node. In this layer, all nodes are linearly transformed to ensure that each node has sufficient representation. A trainable weight matrix W ∈ R shared with all nodes. d×d' is used. Here, d' is the dimension of the transformed node.

[0096] FIG. 8 is a diagram illustrating an attention mechanism for identifying an attention coefficient for each node corresponding to a feature value of patient information according to one embodiment of the present disclosure.

[0097] Referring to FIG. 8, the processor (120) can perform an attention mechanism based on the transformed node. The attention mechanism follows the approach proposed by Bahdanau. Specifically, according to this mechanism, the processor (120) can calculate attention coefficients through a single-layer feedforward neural network.

[0098]

[0099] Here e ij denotes the importance of node j with respect to node i, and A∈R 2d'represents a trainable weight vector. Masked attention coefficients can be added to the mechanism to inject information into the graph structure. Using masked attention coefficients, only e for nodes ij is calculated. j ∈ N i , where N i is a neighbor node of node i (including node i). In addition, adjacent nodes N i To facilitate the comparison of the importance of neighboring nodes N i To facilitate comparison of the importance of e ij can be normalized using the softmax function.

[0100]

[0101] The processor (120) is node H i Normalized attention coefficient α to update the upper node ij A weighted sum operation can be applied to .

[0102] In other words, the processor (120) can identify a second node by performing a linear transformation on the first node and identify the importance of at least one third node adjacent to the second node. Here, the first node may correspond to a different dimension from the dimension corresponding to the second node.

[0103] The processor (120) can normalize the importance of at least one third node to identify an attention coefficient corresponding to each of at least one third node.

[0104] FIG. 9 is a diagram for explaining an operation of updating each node corresponding to a feature value of patient information based on an adjacent node and an attention coefficient of the adjacent node, according to one embodiment of the present disclosure.

[0105] Referring to Figure 9, the processor (120) is a neighboring node N iAnd can perform weighted sum operation. In this process, the processor (120) can apply the following exponential linear (ELU, Exponential Linear Unit) function performed in all nodes. The processor (120) can apply the following exponential linear (ELU) function to the node H i is updated in parallel K times, and the aggregated Hi can be generated using a concatenation operation or an average operation.

[0106] In other words, the processor (120) can identify the fourth node by updating the second node based on at least one third node and an attention coefficient corresponding to each of the third nodes.

[0107] More specifically, the processor (120) can update the second node by performing at least one operation among a chain operation and an average operation based on at least one third node and an attention coefficient corresponding to each of the third nodes, thereby identifying the fourth node.

[0108] According to various embodiments, when there are multiple third nodes, the processor (120) can identify whether the attention coefficient corresponding to each of the third nodes is less than a threshold value. The processor (120) can identify the fourth node by updating the second node only based on the attention coefficients corresponding to the remaining third nodes and the remaining third nodes, excluding the third nodes corresponding to the attention coefficients less than the threshold value. In this case, the third nodes having low attention coefficients and thus importance less than the threshold value are excluded, and the second node is updated as the fourth node, so that unnecessary factors, noise factors, etc. are not included in the node update process.

[0109] The processor (120) may identify the fourth node by updating the second node multiple times based on at least one third node and an attention coefficient corresponding to each of the third nodes.

[0110] The concatenation operation or average operation can be performed using the following mathematical formula 5.

[0111]

[0112]

[0113] Here α k ij is the normalized attention coefficient. W k is the kth weight matrix. The processor (120) can linearly transform Hi into a dimension equal to the dimension d of the multi-head attention. The processor (120) can combine information between layers by applying a skip connection that combines information from multiple layers, and the processor (120) can repeat the above process L times.

[0114] The processor (120) can identify a fifth node by linearly transforming the fourth node, and combine information between at least one single graph attention layer before being input to the multi-head graph attention (220) module by applying a skip connection to the output data of the multi-head graph attention (220) module. Here, the fifth node can correspond to a dimension that is identical to the dimension corresponding to the first node and different from the dimension corresponding to the fourth node.

[0115] The processor (120) can pass all updated nodes from the multi-layer perceptron (230) (MLP, Multi Layer Perceptron) module through a single flattened and fully connected layer. The processor (120) can pass the updated nodes through a single connection layer composed of a function of batch normalization and rectified linear unit (ReLU, Rectified Linear Unit) activation.

[0116] That is, the processor (120) can flatten and align at least one node included in the multi-head graph attention (220) module, and apply the aligned at least one node to a single connection layer including batch normalization and a rectified linear unit (ReLU) activation function.

[0117] Finally, the processor (120) can input patient information into the FT-GAT model to obtain a prediction value for the success rate of spontaneous breathing. For example, the processor (120) can obtain a prediction value for the success rate of spontaneous breathing by calculating a value between 0 and 1 using a sigmoid activation function. Here, the multi-layer perceptron (230) module may include a sigmoid activation function.

[0118] The processor (120) can obtain the patient's spontaneous breathing success rate corresponding to the output value of the multilayer percentron module.

[0119] The FT-GAT design for predicting the success of self-breathing can be implemented through hyperparameters.

[0120] FIG. 10 is a table illustrating hyperparameters required to predict the success rate of an autonomous breathing attempt using the FT-GAT model according to one embodiment of the present disclosure.

[0121] Referring to Figure 10, the hyperparameters used in each process can be confirmed.

[0122] To minimize the impact of imbalanced training data, we adopted the Weighted Binary Cross-Entropy Loss. The loss was weighted based on the imbalance ratio of the training data. Training data samples for the target class (i.e., the self-respiration success rate) were weighted by 0.38. Training was performed for up to 1,000 epochs, and the learning rate was multiplied by 0.5 if the model performance indicator did not increase for 10 epochs. If the model performance indicator did not increase for 40 epochs, model training was terminated, and the model corresponding to the highest performance indicator was used to evaluate the test data.

[0123] FT-GAT was developed using the TensorFlow 2.4.1 library and Python 3.8. Furthermore, the FT-GAT model was trained on a desktop PC using the CUDA 11.0.3 toolkit. The desktop PC's specifications included an Intel Core i7-12,700 @ 2.10GHz central processing unit (CPU) and an NVIDIA GeForce RTX 3090 Ti graphics processing unit (GPU).

[0124] To evaluate the model performance of FT-GAT, five quantitative indices were adopted: accuracy, sensitivity, specificity, F1 score, and area under the curve (AUROC). Here, sensitivity is plotted against 1-specificity on the Y-axis, representing an ROC curve. The evaluation indices were calculated using the following mathematical formulas.

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] Here, TP, FP, TN, and FN represent True Positives, False Positives, True Negatives, and False Negatives, respectively. All values ​​range from 0 to 1, with 1 being the optimal value. In addition, to verify the superiority of FT-GAT, we compared its performance with representative clinical indicators and the most representative clinical indicators. Here, the clinical indicators can be the respiratory rate (RR) and tidal volume ratio (RSBI), and the partial oxygen saturation (PaO2) and oxygen saturation (FiO2) ratio (P / F ratio). The higher the RSBI, the lower the spontaneous breathing success rate. The lower the P / F ratio, the lower the spontaneous breathing success rate.

[0131] Additionally, the performance of FT-GAT can be compared with machine learning (ML) and deep learning (DL) models: categorical boosting (CatBoost), logistic regression, support vector machine (SVM), random forest, light gradient boosting machine (LGBM), extreme gradient boosting (XGBoost) / taptransformer, DANET, and MLP.

[0132] Each model was implemented using Scikit-learn 1.1.2, CatBoost 1.1, and XGBoost 1.6.2, LightGBM 3.3.2, and TensorFlow 2.4.1 libraries.

[0133] Additionally, we used random search, grid search, and Bayesian optimization to set the optimal hyperparameters.

[0134] Statistical analysis was performed to identify and compare patient groups with successful and unsuccessful spontaneous breathing in the dataset and to determine the relationship between each feature and the underlying variable (success / failure of spontaneous breathing).

[0135] Each data type was expressed as mean and standard deviation for continuous variables and as counts and proportions for categorical variables. In addition, each function was expressed as chi-square. We used the test or Fisher's exact test (for categorical features) and the t-test (for continuous features). A P value less than 0.05 was considered statistically significant, and all statistical analyses were performed using R statistical software version 4.1.2 (R Core Team 2021).

[0136] FIG. 11 is a table showing patient information related to predicting the success rate of spontaneous breathing attempts using the FT-GAT model according to one embodiment of the present disclosure.

[0137] Referring to Figure 11, the dataset included 1,041 cases, with more males (67.8%) than females (32.2%). Most patients were aged between 51 and 81 years, and their BMI (22.7) was within the normal range. Furthermore, statistical tests revealed that all features, except for TV 6 mL / kg PBW and HR, were significantly associated with the dependent variable.

[0138] We evaluated the performance of the FT-GAT model using 4-fold cross-validation, which effectively controls model overfitting. Specifically, the dataset was evenly divided into four folds to evenly distribute the number of successful spontaneous breathing attempts. Four validation iterations were then performed, with one fold used for validation (12.5%) and testing (12.5%), and the remaining fold used for training (75%).

[0139] FIG. 12 is a diagram showing experimental data evaluating the performance of the FT-GAT model in predicting the success rate of autonomous breathing attempts according to one embodiment of the present disclosure.

[0140] Referring to Figure 12, the superior performance of the FT-GAT model can be seen based on the results of the 4-fold cross-validation of the FT-GAT model.

[0141] In terms of AUROC, the FT-GAT model achieved a maximum of 0.87. Here, quantitative indicators can be calculated using a confusion matrix.

[0142] FIG. 13 is a table showing experimental data evaluating the performance of the FT-GAT model in predicting the success rate of autonomous breathing attempts according to one embodiment of the present disclosure.

[0143] Referring to Figure 13, CI represents the confidence interval, and the 95% CI for the mean of all folds was calculated using a t distribution with 3 degrees of freedom. In summary, the FT-GAT model achieved an average accuracy of 0.72, a sensitivity of 0.75, a specificity of 0.72, an F1 score of 0.76, and an AUROC of 0.80.

[0144] FIG. 14 is a table showing indicators for explaining performance comparison between the FT-GAT model and other models according to one embodiment of the present disclosure.

[0145] Referring to Figure 14, it can be seen that the indicators of the FT-GAT model have superior values ​​compared to other models.

[0146] To validate each component of FT-GAT, we conducted an ablation study. Specifically, we conducted the ablation study in terms of graph structure and network architecture, and the results are shown in Figure 15. The results were measured using the average area under the curve (AUROC) calculated through four rounds of cross-validation.

[0147] FIG. 15 is a table showing the experimental results for each structure of each node included in the FT-GAT model according to one embodiment of the present disclosure.

[0148] Referring to Figure 15, we can see the experimental results for each node's structure. Regarding graph structure, we compared structured set edges with all edges connecting all nodes, and observed differences in model performance depending on the presence or absence of self-edges. Graphs with self-edges showed better model performance than graphs without self-edges.

[0149] Regarding the network architecture, the FT-GAT model uses a feature tokenization mechanism. We compared its performance with models that do not use a feature tokenization mechanism. Experimental results showed that FT-GAT performs better by converting all input features into embeddings through a feature tokenizer (210).

[0150] FIG. 16 is a diagram showing a comparison of ROC curves of the FT-GAT model and other conventional models according to one embodiment of the present disclosure.

[0151] As shown in Figure 16, the FT-GAT model outperformed all clinical indicators and other models in terms of AUROC. Further details on the performance results for all models are shown in Figure 11.

[0152] FIG. 17 is a diagram showing an average heatmap of normalized attention coefficients according to one embodiment of the present disclosure.

[0153] Referring to Figure 17, the graph composed of set edges with self-edges showed the highest model performance, because the attention coefficient was calculated only at the nodes with strong influence.

[0154] Temporal validation is often used instead of external validation for various reasons, such as IRB approval requirements at other institutions. This involves validating the model against data from new patients included in the same study as the development cohort, but with data sampled earlier or later. In biotechnology research, temporal validation is also performed to assess variables such as changes in patient population, medical device technology, and specialist skills over time.

[0155] Because the dataset consisted of critically ill patients receiving mechanical ventilation between July 2020 and July 2022, temporal cross-validation was performed by dividing the data into the second half of 2020, the first half of 2021, the second half of 2021, and the first half of 2022, based on the start of mechanical ventilation. Each fold was used for validation (50%) and testing (50%), and the model was trained and evaluated anew.

[0156] FIG. 18 is a table showing data representing the performance of an FT-GAT model according to one embodiment of the present disclosure.

[0157] Referring to Figure 18, we can see the results of temporal cross-validation performed considering temporal constraints. FT-GAT achieved slightly lower performance (AUROC: 0.8028) than the performance achieved with standard cross-validation (AUROC: 0.8017), but sufficiently demonstrated generalization ability by evaluating test data from different time points than the training data.

[0158] While mechanical ventilation withdrawal protocols have been widely used in intensive care units (ICUs), deciding whether to attempt spontaneous ventilation or extubation requires a multi-faceted assessment by the attending physician, and artificial intelligence (AI) can assist in this decision-making. Furthermore, researchers have developed predictive indicators, such as the Compliance, Oxygen Generation, Respiration, and Effort (CORE) index, the Compliance, Rate, Oxygenation, and Pressure (CROP) index, the Respiratory Rate, Inspiratory Pressure (RSBI), and peak inspiratory pressure, to identify "readiness for mechanical ventilation" and facilitate spontaneous ventilation attempts. However, no single, robust predictor exists for real-world outcomes. Failure rates for the first spontaneous ventilation attempt range from 25% to 40%, and failure rates for extubation are approximately 10% to 25%, and these rates can vary across institutions and clinicians. Predicting the success of spontaneous ventilation attempts remains challenging for respiratory intensive care unit specialists, and the success rates are relatively low compared to extubation success rates.

[0159] Recent AI research has demonstrated excellent performance in predicting extubation in mechanical ventilation. Pai et al. developed an explainable XGBoost model (AUROC: 0.92) that identified six key feature discriminant points for predicting extubation. Fabregat et al. analyzed heterogeneous data using support vector machines (AUROC: 0.98), and Takanobu et al. used LightGBM to predict reintubation within 72 hours after extubation (AUROC: 0.95). Recently, dim learning methods have also been used in this field. Zeng et al. extracted temporal data from the Medical Information Mart in the Intensive Care Unit (MIMIC-IV) dataset to design an interpretable RNN model (AUROC: 0.83) to dynamically predict extubation failure. Furthermore, Jia et al. used a convolutional neural network model (AUROC: 0.94) using clinically relevant and meaningful features.

[0160] The FT-GAT model has made notable contributions. In the aforementioned studies, the predicted outcome was whether extubation was successful. However, the FT-GAT model predicts spontaneous breathing success, not extubation success.

[0161] Only one study predicting SBT used a machine learning approach, but this study focused on mechanical ventilation mode transitions and considered only t-piece trials to define spontaneous breathing attempts.

[0162] In contrast, the FT-GAT model study included all spontaneous breathing attempts according to the updated definition, and the FT-GAT model showed better performance in terms of AUROC (our model: 0.80 vs. the machine learning model used in the previous study: 0.79).

[0163] Additionally, to increase the generalizability of the predictive model, we included both mixed and trauma intensive care unit patients, and to capture more failure cases, we enrolled patients following a protocol prescribed by an intensivist and patients following a protocol prescribed by a non-respiratory physician (without a protocol) into the dataset.

[0164] To predict the success of spontaneous ventilation attempts, we developed a novel GNN based on a deep learning model. To develop the model, we utilized features available from all intensive care unit patients and considered their relationships. By converting all features (nodes) into embeddings using a feature tokenization mechanism and performing graph analysis using the GAT mechanism, FT-GAT outperformed existing ML and DL models widely used for clinical indicators and tabular data analysis. Specifically, the model achieved an average accuracy of 0.72, a sensitivity of 0.75, a specificity of 0.72, an F1-score of 0.76, and an area under the curve (AUROC) of 0.80. Furthermore, the results of the resection study showed that all features necessary for predicting the success of spontaneous ventilation were successfully converted into a graph structure, and the feature tokenization technique improved model performance.

[0165] In conclusion, we developed a novel GNN-based deep learning model, the FT-GAT model, to predict the success of spontaneous ventilation in patients. It demonstrated excellent performance, with an average accuracy of 0.72, sensitivity of 0.75, specification accuracy of 0.72, F1 score of 0.76, and area under the curve of 0.80.

[0166] According to one embodiment, the method according to the various embodiments disclosed in the present document may be provided as included in a computer program product. The computer program product may be traded as a product between a seller and a buyer. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., compact disc read only memory (CD-ROM)), or may be distributed online (e.g., downloaded or uploaded) via an application store (e.g., Play Store™) or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., downloadable app) may be temporarily stored or temporarily generated in a machine-readable storage medium, such as a memory (110) of a manufacturer's server, an application store's server, or a relay server.

[0167] The artificial intelligence-related function according to the present disclosure is operated through the processor (120) and memory (110) of the electronic device (100).

[0168] The processor (120) may be composed of one or more processors (120). At this time, the one or more processors (120) may include at least one of a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), and an NPU (Neural Processing Unit), but is not limited to the examples of the processors (120) described above.

[0169] The CPU is a general-purpose processor (120) capable of performing not only general calculations but also artificial intelligence calculations. Its multi-layer cache structure allows for the efficient execution of complex programs. The CPU is advantageous in a serial processing method, enabling organic linking of previous and subsequent calculation results through sequential calculations. The general-purpose processor (120) is not limited to the examples described above, except in cases where it is specifically referred to as a CPU.

[0170] A GPU is a processor (120) for large-scale operations such as floating point operations used in graphic processing, and can perform large-scale operations in parallel by integrating a large number of cores. In particular, a GPU may be advantageous compared to a CPU in parallel processing methods such as convolution operations. In addition, a GPU may be used as a co-processor (120) to supplement the functions of a CPU. The processor (120) for large-scale operations is not limited to the examples described above, except in cases where it is specified as a GPU as described above.

[0171] An NPU is a processor (120) specialized in artificial intelligence operations using an artificial neural network, and each layer constituting the artificial neural network can be implemented with hardware (e.g., silicon). At this time, since the NPU is designed specifically according to the required specifications of the company, it has a lower degree of freedom compared to a CPU or GPU, but it can efficiently process the artificial intelligence operations requested by the company. Meanwhile, as a processor (120) specialized in artificial intelligence operations, the NPU can be implemented in various forms such as a TPU (Tensor Processing Unit), an IPU (Intelligence Processing Unit), a VPU (Vision processing unit), etc. The artificial intelligence processor (120) is not limited to the above-described examples, except in cases where it is specified as the above-described NPU.

[0172] Additionally, one or more processors (120) may be implemented as a SoC (System on Chip). In this case, the SoC may further include, in addition to one or more processors (120), a memory (110), and a network interface such as a bus for data communication between the processor (120) and the memory (110).

[0173] When a plurality of processors (120) are included in a SoC (System on Chip) included in an electronic device (100), the electronic device (100) may perform operations related to artificial intelligence (e.g., operations related to learning or inference of an artificial intelligence model) by using some of the processors (120) among the plurality of processors (120). For example, the electronic device (100) may perform operations related to artificial intelligence by using at least one of a GPU, an NPU, a VPU, a TPU, and a hardware accelerator specialized in artificial intelligence operations such as convolution operations and matrix multiplication operations among the plurality of processors (120). However, this is merely an example, and it is of course possible to process operations related to artificial intelligence by using a CPU or a general-purpose processor (120).

[0174] In addition, the electronic device (100) can perform operations related to functions related to artificial intelligence by utilizing multiple cores (e.g., dual cores, quad cores, etc.) included in one processor (120). In particular, the electronic device (100) can perform artificial intelligence operations such as convolution operations, matrix multiplication operations, etc. in parallel by utilizing multiple cores included in the processor (120).

[0175] One or more processors (120) are controlled to process input data according to predefined operation rules or artificial intelligence models stored in the memory (110). The predefined operation rules or artificial intelligence models are characterized by being created through learning.

[0176] Here, "created through learning" means that a predefined set of behavioral rules or an AI model with desired characteristics is created by applying a learning algorithm to a large number of learning data. This learning may be performed on the device itself, where the AI ​​according to the present disclosure is implemented, or through a separate server / system.

[0177] An artificial intelligence model may be composed of multiple neural network layers. At least one layer has at least one weight value and performs its operation through the operation result of the previous layer and at least one defined operation. Examples of neural networks include a convolutional neural network (CNN), a deep neural network (DNN), a recurrent neural network (RNN), a restricted boltzmann machine (RBM), a deep belief network (DBN), a bidirectional recurrent deep neural network (BRDNN), deep Q-networks, and a transformer. The neural networks in the present disclosure are not limited to the above-described examples unless otherwise specified.

[0178] A learning algorithm is a method for training a target device (e.g., a robot) using a large amount of learning data, enabling the target device to make decisions or predictions on its own. Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. Unless otherwise specified, the learning algorithms in this disclosure are not limited to the aforementioned examples.

[0179] Although the preferred embodiments of the present disclosure have been illustrated and described above, the present disclosure is not limited to the specific embodiments described above, and various modifications may be made by a person having ordinary skill in the art to which the present disclosure pertains without departing from the gist of the present disclosure as claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present disclosure.

Claims

1. In electronic devices, A memory storing an FT-GAT (Feature Tokenizer - Graph Attention Network) model for obtaining a prediction value of a patient's spontaneous breathing success rate based on patient information; and A processor connected to the memory and configured to input at least one feature value of the patient into the FT-GAT model to identify a prediction value of the patient's spontaneous breathing success rate; The above FT-GAT model, An electronic device comprising a feature tokenizer module that converts feature values ​​corresponding to the above patient information into embedding values ​​in a latent space, a multi-head graph attention module that updates nodes corresponding to the feature values, and a multi-layer perceptron (MLP) module that applies the updated nodes to output a prediction value of the patient's spontaneous breathing success rate corresponding to the embedding values.

2. In paragraph 1, The above processor, The embedding value is obtained by multiplying the feature value corresponding to the above patient information by a weight and adding a bias, The above feature values ​​are, An electronic device, comprising at least one of a category corresponding to the patient information and a quantitative value corresponding to the patient information.

3. In paragraph 1, The multi-head graph attention module is, At least one single graph attention layer including a first node corresponding to the above patient information, The above processor, Identify the second node by performing a linear transformation on the first node, Identify the importance of at least one third node adjacent to the second node, Normalizing the importance of at least one third node to identify an attention coefficient corresponding to each of the at least one third node, An electronic device that identifies a fourth node by updating the second node based on the at least one third node and the attention coefficient corresponding to each of the third nodes.

4. In paragraph 3, The above processor, An electronic device that updates the second node by performing at least one operation among a chain operation and an average operation based on the at least one third node and the attention coefficient corresponding to each of the third nodes, thereby identifying the fourth node.

5. In paragraph 3, The above processor, An electronic device that identifies the fourth node by updating the second node multiple times based on the at least one third node and the attention coefficient corresponding to each of the third nodes.

6. In paragraph 3, The above processor, Identify the fifth node by linearly transforming the fourth node, An electronic device that combines information between at least one single graph attention layer before inputting the multi-head graph attention module by applying a skip connection to the output data of the multi-head graph attention module.

7. In paragraph 5, The above first node, Corresponds to a dimension different from the dimension corresponding to the second node above, Corresponds to a dimension different from the dimension corresponding to the above fourth node, An electronic device corresponding to the same dimension as that corresponding to the fifth node.

8. In paragraph 1, The above processor, Flatten and align at least one node included in the above multi-head graph attention module, An electronic device that applies at least one of the above-mentioned aligned nodes to a single-connected layer including batch normalization and a rectified linear unit (ReLU) activation function.

9. In paragraph 1, The above multilayer perceptron module, Includes a sigmoid activation function, The above processor, An electronic device that obtains the patient's spontaneous breathing success rate corresponding to the output value of the multilayer percentron module.

10. A method for controlling an electronic device including an FT-GAT (Feature Tokenizer - Graph Attention Network) model, A step of inputting feature values ​​corresponding to patient information into a tokenizer module to identify embedding values ​​in the latent space; A step of controlling a multi-head graph attention module to update a node corresponding to the above feature value; and A control method, comprising: a step of inputting the embedding value into the multi-layer perceptron module to which the updated node is applied to obtain a prediction value of the patient's spontaneous breathing success rate.

11. A non-transitory computer-readable recording medium storing at least one instruction that is executed by a processor of an electronic device to cause the electronic device to perform the control method of claim 10.

Images