Artificial intelligence-based biomarker selection device and method

Abstract

The embodiment relates to an AI-based biomarker selection device and method, which includes an acquisition unit for acquiring biodata, an encoder for receiving the biodata and calculating a numerical vector including one or more elements that are biomarker candidates, and a biomarker screening unit for screening biomarkers from the one or more elements of the numerical vector. The device and method of the embodiment can be used to discover various biomarkers that can be used for drug discovery and prognosis prediction, such as development of new cardiovascular drugs. As a result, the device and method can be effectively used in pharmaceutical platforms for new drug development and research platforms for precision medicine, disease diagnosis, treatment optimization, etc.

Description

[Technical field]

[0001] The present specification relates to an artificial intelligence-based biomarker selection device and method, and more particularly, to an artificial intelligence-based biomarker selection device and method capable of developing patient-derived information-based companion diagnostic biomarkers for new drug development, precision medicine, disease diagnosis, and treatment optimization.

[0002] <CROSS REFERENCE TO RELATED APPLICATIONS> This application claims priority to Korean Provisional Patent Application No. 10-2022-0013491, filed on January 28, 2022, the entire contents of which are incorporated by reference into this application. [Background technology]

[0003] A companion diagnostic is a test used to determine whether a patient with a particular disease can be treated with a particular drug through the evaluation of a biomarker.

[0004] Here, biomarkers include predictive biomarkers that can distinguish between responders and non-responders to a particular drug, and prognostic biomarkers that can predict the prognosis of a disease and indicate the progression of a disease or whether a group requires treatment.

[0005] Existing traditional biomarkers generally involve the selection of biological markers using techniques such as polymerase chain reaction (PCR), in situ hybridization (ISH), next-generation sequencing (NGS), and immunohistochemical staining (IHC).

[0006] However, most companion diagnostics are only used in conjunction with major diseases such as cancer and drugs to treat them, and biomarkers are often not available for other diseases or drugs.

[0007] For example, in the case of cardiovascular drugs, which are widely prescribed drugs and taken by the majority of patients for life, there is a need for companion diagnostics, but there are no effective biomarkers.

[0008] One of the reasons why there are no effective biomarkers for some diseases or no attempts to find such biomarkers are made is that it is extremely difficult, time-consuming, and costly to select biomarkers using traditional methods as described above. Summary of the Invention [Problem to be solved by the invention]

[0009] In one aspect, the embodiments of the present application relate to an artificial intelligence-based biomarker selection device and method capable of developing companion diagnostic biomarkers that can be utilized in new drug development, precision medicine, disease diagnosis, and treatment optimization for various diseases, including diseases for which no existing biomarkers exist, such as cardiovascular disease. [Means for solving the problem]

[0010] In an embodiment of the present application, there is provided an artificial intelligence based biomarker selection method, the method including: an encoding step of encoding biodata to calculate a numerical vector including one or more elements that are biomarker candidates; and a biomarker screening step of screening a biomarker from the one or more elements of the numerical vector; and a computer readable recording medium having a program for executing the method recorded thereon, or a computer program stored in the computer readable recording medium.

[0011] In addition, an embodiment of the present application provides an artificial intelligence-based biomarker selection device, which includes an acquisition unit that acquires biodata; an encoder that receives the biodata and calculates a numerical vector including one or more elements that are biomarker candidates; and a biomarker screening unit that screens biomarkers from the one or more elements of the numerical vector. Effect of the Invention

[0012] In one aspect, the embodiments of the present application can be utilized to discover various biomarkers that can be used for drug discovery, such as the development of new cardiovascular drugs, prognosis prediction, etc. This can be effectively utilized in pharmaceutical platforms for new drug development, and research platforms for precision medicine, disease diagnosis, treatment optimization, etc.

[0013] The effects of the present application are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description of the claims. [Brief description of the drawings]

[0014] In order to more clearly describe the exemplary embodiments of the present application, the drawings necessary for the description of the embodiments are briefly introduced below. It should be understood that the following drawings are only intended to describe the embodiments of the present specification, and are not intended to be limiting. In addition, for the sake of clarity, some elements may be shown in the following drawings to which various modifications such as exaggeration and omission have been applied. [Figure 1] FIG. 1 is a schematic diagram showing the configuration of an artificial intelligence-based biomarker selection device according to one aspect of the present application. [Figure 2a-2b]FIG. 2a is a schematic diagram showing the encoder architecture of an artificial intelligence-based biomarker selection device in one embodiment of the present application, and FIG. 2b is a schematic diagram illustrating the process of finding matching pairs and non-matching pairs in the common space of FIG. 2a. [Figure 3a-3c] Fig. 3a is a schematic diagram showing a numerical vector including one or more elements that are biomarker candidate groups in one embodiment of the present application, Fig. 3b is an explanatory diagram explaining the structure and role of the numerical vector, and Fig. 3c is a schematic diagram explaining a process for improving the embedding quality of each element of the numerical vector in one embodiment of the present application. [Figure 4a-4b] In one embodiment of the present application, a schematic diagram is shown illustrating selection by clustering in a biomarker screening process, in which FIG. 4a illustrates each cluster, and FIG. 4b illustrates a numerical array (matrix) constructed by extracting m candidate biomarkers from n subjects (or samples). [Figure 5a-5b] In one embodiment of the present application, a schematic diagram of a process for generating effect modifier biomarker m that indicates the effect of a drug or treatment in a biomarker screening process and a biomarker n that is unrelated to effect modifier using an artificial neural network is shown in FIG. 5a, which shows a case where a numerical vector Z passes through one neural network, and FIG. 5b, which shows a case where the numerical vector Z passes through different neural networks. [Figure 6] FIG. 1 is a schematic diagram illustrating a process for measuring effect modification in a biomarker screening process in one embodiment of the present application. [Figures 7a-7c]In one embodiment of the present application, a schematic diagram showing a process of finding biomarker attributes in a biomarker screening process is shown. Fig. 7a is a schematic diagram summarizing the characteristics of a numeric vector Z, in which m biomarker candidates belonging to the numeric vector Z in an m x n matrix each have n attribute vectors X. Fig. 7b is a diagram illustrating a case where all biomarker candidates are divided into elements of a morphological numeric vector Z', a clinical numeric vector Z'', and a phenotypic numeric vector Z''', and Fig. 7c is a diagram illustrating obtaining an attribute vector of a morphological numeric vector based on similarity. [Figure 8a-8b] In one embodiment of the present application, a schematic diagram for explaining an artificial intelligence-based biomarker selection method is shown in Fig. 8. Fig. 8a is a diagram for explaining each step of the method, and Fig. 8b is a schematic diagram showing a numerical vector, which is a group of biomarker candidates obtained in the process of Fig. 8a, particularly when electrocardiogram data is used, and a biomarker obtained by selecting a biomarker therefrom. [Figure 9a-9c] According to one embodiment of the present invention, a new electrocardiogram biomarker was developed to predict the risk of death of a COVID-19 infected patient using a numerical vector Z extracted from an existing electrocardiogram analysis encoder, and the experimental data was evaluated for its sensitivity, specificity, etc.

[0015] Some embodiments of the present application will be described in detail below with reference to the drawings. When referring to the components in each drawing, the same components may be referred to by the same reference numerals as much as possible even if they are shown in different drawings. In addition, when describing the present embodiment, if it is determined that a detailed description of related publicly known configurations or functions may obscure the gist of the present technical idea, the detailed description may be omitted.

[0016] Definition of Terms In this specification, when "comprises," "has," "consists of," etc. are used, other parts may be added unless "only" is used. When an element is expressed in the singular, it may include a plural unless otherwise expressly stated.

[0017] In addition, in describing components of this application, terms such as first, second, A, B, (a), (b), etc. may be used. Unless otherwise specified, these terms are merely used to distinguish the components from other components, and are not intended to limit the essence, order, procedure, number, etc. of the components.

[0018] As used herein, "learning" is a term that refers to performing machine learning through procedural computing.

[0019] As used herein, network refers to a neural network of machine learning algorithms or models.

[0020] In this specification, terms such as "unit," "module," "apparatus," or "system" are intended to refer to a combination of not only hardware but also software driven by the hardware. For example, hardware may be a data processing device including a central processing unit (CPU), a graphic processing unit (GPU), or other processor. In addition, software may refer to a running process, an object, an executable, a thread of execution, a program, etc.

[0021] As used herein, biomarker refers to a digital biomarker.

[0022] As used herein, a biomarker candidate refers to a numerical vector or elements of a numerical vector before being selected as a biomarker.

[0023] In this specification, biodata refers to information derived from a patient that can be used for new drug development, precision medicine, disease diagnosis, treatment optimization, etc., and may be, for example, a biosignal such as an electrocardiogram, but is not limited to this.

[0024] As used herein, modality may refer to a formal characteristic, such as a signal or image of a waveform of a biosignal.

[0025] In this specification, a platform or platform system may mean one that at least partially includes the artificial intelligence-based biomarker selection device of the present invention. For example, a pharmaceutical platform for new drug development or a research platform for various precision medicines, disease diagnosis, treatment optimization, etc. may at least partially include the artificial intelligence-based biomarker selection device of the present invention, or may be the device itself.

[0026] As used herein, common space refers to a shared coordinate system in which similarities between numerical vectors can be expressed and compared, and refers to a dimensional space onto which numerical vectors are projected in order to find pairs between the numerical vectors.

[0027] The various data mining techniques mentioned herein, such as clustering, dimensionality reduction, etc., are intended to encompass all data mining methodologies applicable based on error vectors and are not limited to the specific examples presented.

[0028] Description of exemplary embodiments FIG. 1 is a schematic diagram showing the configuration of an artificial intelligence-based biomarker selection device according to one aspect of the present application.

[0029] Referring to FIG. 1, an artificial intelligence-based biomarker selection device may include an acquisition unit 10 that acquires biodata, an encoder 20 that receives input of the biodata from the acquisition unit 10 and calculates a numerical vector including one or more elements that are biomarker candidates, and a biomarker screening unit 30 that screens biomarkers from one or more elements of the numerical vector.

[0030] The acquisition unit 10 acquires biodata, which is information derived from one or more patients, from an acquisition device (not shown).

[0031] The biodata may be two or more pieces of biodata having different or the same modality. In one example, the biodata may include at least one of waveform signal data and image data. When the biodata is an electrocardiogram, the biodata as input data may include at least one of an electrocardiogram signal, which is a one-dimensional single-channel or multi-channel signal, and an electrocardiogram image in which the electrocardiogram signal is illustrated on a two-dimensional plane.

[0032] In one example, the signal is in the form of a two-dimensional array, for example, C×T (number of input leads (channels)×number of measurements for each channel), and the image may be an image including all lead channels, or a patch image for each lead channel cropped for each lead channel, and may be a black and white image of one or more lead channels, or a three-dimensional array of C×W×H (number of channels×number of horizontal pixels×number of vertical pixels) having three channels, R (ReD), G (Green), and B (Blue).

[0033] Meanwhile, the biodata may be any kind of information derived from a patient. In an exemplary embodiment, as described below, it may be representatively electrocardiography (ECG) data, but is not limited thereto, and may include various time-series biosignals such as electroencephalogram (EEG), electromyogram (EMG), electrooculography (EOG), etc. In addition, it may have any modality such as waves, images, sounds, and other forms, but is not limited to a specific modality. Furthermore, the biodata may also include various basic patient data such as the patient's gender, age, medical history, etc.

[0034] The encoder 20 may be trained through supervised or self-supervised learning to provide a numerical vector having one or more elements that are candidate biomarkers, and may include various neural network structures adapted to provide the numerical vector (hereinafter, sometimes referred to as the numerical vector Z).

[0035] In one example, the encoder 20 may include at least one of a clinical encoder and a clinical semantics and morphology encoder.

[0036] The clinical encoder may be configured to receive biodata and compute a first numeric vector including one or more elements associated with clinical meaning through supervised learning associated with a task having clinical meaning.

[0037] The morphological encoder may be configured to receive biodata and calculate a second numeric vector including one or more elements associated with morphological features of the biodata, and may be trained by unsupervised learning.

[0038] The first numeric vector Z1 that has been subjected to the clinical encoder and the second numeric vector Z2 that has been subjected to the morphological encoder may be transformed for the same biodata input through a stochastic process, such as test-time random augmentation, Monte-Carlo dropput (or dropconnect or a similar known technique), Bayesian layer, etc.

[0039] In one example, the encoder 20 may further provide a new numerical vector by concatenating the first numerical vector Z1 and the second numerical vector Z2. For this purpose, the neural network of the encoder 20 may be trained by multi-task learning in which supervised learning and unsupervised learning are integrated.

[0040] In one example, at least one of the mean, range, and standard deviation (SD) of some or all of one or more elements of each of the numerical vectors Z1 and Z2 repeatedly calculated from the encoder 20 may be included in the biomarker candidates.

[0041] Furthermore, the encoder 20 may further project the calculated numerical vectors onto a common space and use the projected vectors as training data to find pairs based on the similarity between the numerical vectors.

[0042] For example, the encoder 20 may be trained in a manner that projects the calculated numerical vectors onto a common space using another artificial neural network that is trained together, and finds positive pairs based on the similarity between the numerical vectors.

[0043] Here, a positive pair can be said to be a matching pair of the same subject within a defined time window frame.

[0044] And / or, the encoder 20 may be trained in a manner that projects the calculated numerical vectors onto a common space and finds negative pairs based on the similarity between the numerical vectors.

[0045] Here, a negative pair may be a pair of the same subject outside a defined time window frame or a pair between different subjects, which may also be referred to as a non-matching pair.

[0046] Here, the determined time window frame may be set, for example, to a certain number of minutes, a certain number of hours, a certain number of days, etc., and the time window size may be adjusted according to the state of the subject according to a main projection or an anchor (i.e., a main numerical vector of a pair) projected onto a common space. That is, the determined time window frame may have an adaptive time window size. For example, when electrocardiogram data obtained from a patient in an emergency room is input and a pair is found by an encoder, the time window size may be narrow, and when electrocardiogram data obtained from an outpatient is used, the time window size may be wide.

[0047] Additionally, the encoder 20 may further use the calculated numerical vectors for supervised learning task training to have clinical meaning.

[0048] FIG. 2a is a schematic diagram showing the encoder architecture of an artificial intelligence-based biomarker selection device in one embodiment of the present application, and FIG. 2b is a schematic diagram illustrating the process of finding matching pairs (positive pairs) and non-matching pairs (negative pairs) in the common space of FIG. 2a.

[0049] As shown in Figures 2a and 2b, the encoder 20 may have a structure having a bottom layer 21 and a common head 23 for one or more inspection methods, i.e., modalities, and may be trained in a manner that projects a numerical vector calculated by the common head 23 onto a common space 25 to find matching pairs and / or non-matching pairs.

[0050] The bottom layer 21 may be a layer for each modality, and receives input of each biodata and outputs a series of vectors that are input to a common head 23 .

[0051] In one example, as part of the pre-processing process when multiple signals and / or images are provided to the bottom layer, a patching process and position embedding may be applied to the original inputs, where the position embedding may be performed according to T (time) for signals and H (height) and W (width) for images.

[0052] Here, the bottom layer 21 corresponds to various artificial neural networks having structures suitable for each modality. These include multi-layer perceptron (MLP, hereinafter sometimes referred to as MLP), CNN (Convolution Neural Network, hereinafter sometimes referred to as CNN), transformer, RNN (Recurrent Neural Network, hereinafter sometimes referred to as RNN), and various modifications thereof, and are not limited to a specific structure of the artificial neural network. The common head 23 layer may have, for example, a transformer neural network structure and / or an MLP structure.

[0053] The vectors that have passed through the common head 23 layer may be projected onto a common space, where they may be trained in a way that finds positive pairs and / or negative pairs based on the similarity between the numerical vectors.

[0054] In one example, the encoder 20 may be trained to find pairs and further trained to have clinical meaning by performing supervised learning tasks, i.e., in addition to matching biodata such as electrocardiograms of the same person within a certain time range, the encoder 20 may be trained to predict the patient's condition and diagnosis at the time of examination, thereby performing specific tasks together so that the encoding result has clinical meaning.

[0055] In one example, the supervised task may be a prediction-related task, such as a task of simultaneously or sequentially predicting one or more of binary, multi-category, and numerical values, such as the presence or absence of a disease in a patient (binary), a classification task (multi-category), or a task of predicting a specific measurement value (numerical).

[0056] In one example, when training to find pairs in the common space 25, the loss function may include a Noise-Contrastive-Estimation (NCE) loss.

[0057] In another example, the loss function may use MIL Multiple-Instance-Learning-NCE (MIL NCE) loss in addition to the Noise-Contrastive-Estimation (NCE) loss, and / or may further include a supervised task loss if a supervised task is performed. In this case, the total loss may be expressed as the following [Equation 1]. [Formula 1] Total loss = NCE loss + λMIL NCE loss + τsupervised task loss

[0058] In the above [Equation 1], λ and τ are hyperparameters that may be adjusted before use, and may each be 0.

[0059] Meanwhile, the positive pair or negative pair may include one or more of a signal and signal pair, an image and image pair, and a signal and image pair. Here, a signal refers to one-dimensional time series data of a single channel or multiple channels, and an image refers to a waveform image that represents the same on a two-dimensional plane in various formats.

[0060] In one example, the signals and / or images may be data augmented, which is the application of various transformations to the original input data, and may use any known data augmentation technique, such as rotation, flip, channel drop, cut out, cut mix, etc.

[0061] When training positive ECG pairs of the same type (e.g., signal-signal, image-image) to be located close to each other in a common space, the first and second encoders can be trained in the same or similar manner. Specifically, a specific encoder (first encoder) capable of processing an input signal and a second encoder having the same structure and parameters as the first encoder (weight-sharing), or a second encoder (momentum encoder) created by applying a weighted average to the most recent parameter values ​​of the first encoder during the training process, can be used.

[0062] In one example, the encoder 20 includes a first encoder for receiving a data-augmented first biodata set and a second encoder for receiving a data-augmented second biodata set, which may be random augmentation of the same electrocardiogram or signals or images within a defined time range of the same patient (subject).

[0063] Here, the first encoder may be any neural network that converts input values ​​into a numerical vector, and the second encoder may be a momentum encoder that shares the same weight as the first encoder or averages the temporal changes of the first encoder.

[0064] Here, the first encoder may or may not further include an MLP layer, which may be used to convert the calculated numerical vector into a new numerical vector.

[0065] Also, the second encoder may or may not further perform clustering to convert the calculated numeric vector into a limited number of embedded numeric vectors. The clustering is a task of selecting an embedded numeric vector that is closest to the calculated numeric vector, and serves to replace the original numeric vector output with the closest embedded numeric vector (representative value).

[0066] A gradient flow process may be performed on the numerical vector that has undergone this process, in which the parameters of the encoder neural network are adjusted using a similarity function and / or a dissimilarity function as a loss function. For example, when a dissimilarity function is used, the gradient flow may be sent to both the first encoder and the second encoder, or to the first encoder, to adjust the parameters of each encoder.

[0067] Here, a similarity function may be used between positive pairs, and a dissimilarity function may be used between negative pairs.

[0068] Additionally, the above-mentioned supervised tasks may be further performed to impart clinical meaning to the gradient flow process.

[0069] The total loss when using such a similarity function and / or dissimilarity function and further performing a supervised task may be expressed as follows: [Formula 2] Total loss = γ similarity loss + λ dissimilarity loss + τ supervised task loss

[0070] In the above [Equation 2], γ, λ, and τ are hyperparameters that may be adjusted before use, and each may be 0.

[0071] FIG. 3a is a schematic diagram showing a numerical vector including one or more elements that are a group of biomarker candidates in one embodiment of the present application, and FIG. 3b is an explanatory diagram explaining the structure and role of the numerical vector.

[0072] As shown in Fig. 3a, the numeric vector Z obtained through the above encoder 20 may include one or more elements. Here, each of the elements is a biomarker candidate. As will be described later, each of the elements may be labeled and stored with an attribute.

[0073] Here, at least some or all of the elements may be combined linearly or nonlinearly to generate new biomarker candidates or biomarkers, and at least some of the elements may be combined after transformation, e.g., polynomial, sigmoid, binary, log sqrt, etc., or further transformations such as those described above may be applied after the combination.

[0074] In one embodiment, at least a portion of one or more elements of the numerical vector may be combined with known clinical biomarkers through linear or nonlinear transformation and used as biomarker candidates. In this process, a machine learning algorithm including a stepwise logistic regression using AIC may be used. The clinical biomarkers may include at least one of age, sex, SOFA score, Troponin I, and Pro-BNP value.

[0075] In one example, as described above, at least one of the mean, range, and standard deviation (SD) of one or more elements of multiple numerical vectors calculated using the same input data may be included in the biomarker candidates.

[0076] As shown in FIG. 3b, the numerical vector Z may include a prediction vector, which is an output value associated with a specific task through supervised learning, and may also include an embedding vector (first embedding vector) obtained (extracted) at an intermediate stage before obtaining the prediction vector.

[0077] The numerical vector Z may also include an embedding vector (second embedding vector) obtained through self-supervised learning or unsupervised learning.

[0078] The numeric vector Z may also include various basic patient data such as the presence or absence of a patient's medical history or various measurement values.

[0079] Among the biovector candidates, the patient data and a prediction vector obtained through supervised learning may be referred to as an explainable biomarker candidate. Such a prediction vector may include, for example, a probability value of the presence or absence of a specific condition or a regression value for a measured value. Such an explainable biomarker candidate may be a clinical biomarker having clinical meaning. Such a clinical biomarker may be effectively used, for example, to select patients corresponding to the indications of a drug. In addition, a pattern of drug effect predicted by a drug structure may be matched with a known biomarker profile for use.

[0080] Meanwhile, the second embedding vector, which is an embedding vector obtained through self-supervised learning (or unsupervised learning), may be a black box type biomarker candidate and a morphological biomarker. Although the association with the clinical characteristics of the patient cannot be known at first, the more research and analysis data on the correlation between these and the clinical characteristics is accumulated, the more explainable they become. If they become explainable to a certain level, they may be included in the explainable biomarker candidate.

[0081] Meanwhile, in the case of the first embedding vector calculated during the supervised learning process, since it corresponds to a calculation process prior to the final biomarker prediction, it may be classified as a black box type biomarker candidate whose clinical meaning is still difficult to explain. Even in this case, it becomes explainable as more research and analysis data on the correlation with clinical characteristics is accumulated, and if it becomes explainable to a certain level, it may be included in the explainable biomarker candidate.

[0082] Meanwhile, in one example, the encoder 20 may use at least one of a Bayesian layer and a KL loss function to reduce correlation between multiple elements.

[0083] FIG. 3c is a schematic diagram illustrating a process for improving the embedding quality of each element of a numerical vector in an embodiment of the present application. For this purpose, a Bayesian Layer may be added to the end of the encoder that creates the embedding vector, and the layer may output a μ (mean) vector and a σ (variance) vector that define a multivariate normal distribution used to extract the embedding vector elements. Here, in order to approximate the multivariate normal distribution to a unit Gaussian (normal) distribution, the Kullback-Leibler (KL) divergence between two multivariate distributions N(μ, σ) and N(0, 1) is calculated, and the KL divergence is minimized by reflecting it in the total loss calculation formula. [Formula 3] KL(N(μ,σ)|N(0,1))

[0084] In the above formula, KL(●|●) represents the Kullback-Leibler divergence, and N(●) represents a normal distribution. By training the encoder in this way, the correlation between the embedding vector elements can be reduced, and the unique features of each element can be made more prominent.

[0085] Next, biomarker screening will be described. The screening unit 30 may screen biovectors from the group of biovector candidates through a regression or decision tree that selects variables in a variety of ways, clustering, dimension reduction, supervised learning, etc.

[0086] In one example, the screening unit 30 may select one or more elements that are candidate biomarkers as a biomarker after grouping them.

[0087] In the case of a group of candidate biomarkers that have undergone the above-mentioned supervised learning prediction, they can be analyzed on their own and can be screened as biomarkers according to the prediction task for which they were trained.

[0088] On the other hand, in the case of a numerical vector Z that is not related to supervised learning, the vectors may be grouped and then selected using one or more processes, such as a clustering process based on similarity and a selection process using a regression equation, either alone or in combination.

[0089] FIG. 4 is a schematic diagram illustrating the selection by clustering in the biomarker screening process in one embodiment of the present application, in which FIG. 4a illustrates each cluster, and FIG. 4b illustrates a numerical array (matrix) constructed by extracting m candidate biomarkers from n subjects (or samples).

[0090] When a clustering technique is applied to the array shown in Fig. 4b to assign m candidate biomarkers to groups less than m, each group corresponds to each cluster shown in Fig. 4a. In this case, the candidate biomarkers belonging to one cluster can be estimated to be biomarkers that show similar tendencies, and if there is a known explainable biomarker in the cluster (e.g., heart failure element or hyperkalemia element in Fig. 4a), it can be estimated based on this that other biomarkers also show similar tendencies. The clustering algorithm used here can be any clustering method that can be generally used in data mining.

[0091] Meanwhile, in one example, the screening unit 30 may store one or more elements of the numerical vector that is the biomarker candidate together with attribute information.

[0092] For example, even if a group of numerical vector elements that are morphological biomarker candidates has empirical analysis, this may be tagged as an attribute of each element and used.

[0093] Meanwhile, in one embodiment, when the screening unit 30 selects some of the elements as biomarkers, it may first preset a value to be predicted, label each of the plurality of biodata, and select the label as a biomarker through a regression equation.

[0094] The regression equation may be a logistic regression equation when predicting a binary event, or a regression equation for predicting a general value. Such a regression equation (R1) can be expressed as the following [Equation 4]. [Formula 4] Q1=af(Z)+c

[0095] In the above formula, f can be a linear or nonlinear transformation function, including the raw identity function, a is the effect size coefficient, and c is the intercept.

[0096] When a p-value indicating statistical significance using such a regression equation [or a value obtained by correcting the p-value using a method such as Bonferroni correction] is less than a preset value, and / or when the absolute value of the effect size coefficient a is equal to or greater than a preset value, the corresponding biomarker candidate Z may be selected as a biomarker.

[0097] The numeric vector element values ​​(z1, z2, z3,,,) selected in this manner may be further screened as follows.

[0098] That is, the selected numerical vector may be subjected to a lasso regression method or an elastic net regression method to select a numerical vector whose coefficient value is not 0 as a biomarker.

[0099] On the other hand, while the regression method described above may be primarily used to predict specific values ​​or events, an effect modification method may be performed to predict the effect of exposure to a drug or the effect of exposure to a specific treatment, in which exposure to a drug, treatment, etc. may be included in the regression equation.

[0100] In this case, the regression equation may be, for example, the following [Equation 5]. [Formula 5] Q2=a·z n ·Exposure+b·z n +c·Exposure+d

[0101] The exposure may be binary (ie, 0 or 1) or other numeric value for exposure to a drug effect or exposure to a particular treatment.

[0102] The regression equation Q2 is used to evaluate whether the effect modification of a biomarker candidate Z on a target of an exposure is statistically significant. n In other words, the effect size is the absolute value of a, and it is possible to check whether the interaction between the biomarker and the exposure is statistically significant or not based on the effect size value or the p-value (used to evaluate statistical significance) of the effect size value [or the value corrected using a correction method such as the Bonferroni correction]. In other words, the interaction term in the above formula, a·z nFor an exposure, if the p-value (or a value obtained by correcting the p-value using a method such as the Bonferroni correction) generally reported in regression analysis is less than a predetermined value, and / or if the absolute value of the coefficient a, which is the effect size, is equal to or greater than a predetermined value, the corresponding biomarker candidate can be selected as an effect-modifying biomarker for the specific exposure.

[0103] Biomarker candidate z in the above [Formula 5] n The value may be used after passing it through a linear or nonlinear transformation function including an identity function, as in [Equation 4]. That is, the regression equation Q3 of [Equation 6] below may be used. [Formula 6] Q3=Exposure·(a 1· f1(z1)+a 2· f2(z2)+...+a n· f n (z n ))+b 1· g1(z1)+b 2· g2(z2)+...+b n· g n (z n )+c · Exposure +d

[0104] In the above formula, the f and g functions may be linear or nonlinear transformation functions including an identity function. The regression analysis method used here may be the Lasso regression method or the Elastic Net regression method as described above, or the AIC- and BIC-based regression method.

[0105] Here, the coefficients related to weight decay in the lasso regression method and the elastic net regression method may be adjusted using hyperparameters.

[0106] AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) based regression methods are methods for selecting a subset of predictor variables in a multiple regression equation, and the set of variables with the lowest AIC or BIC value is used in the regression equation, thus reducing the number of variables and simplifying the model while maintaining the regression model fit to the data.

[0107] Through the above process, z1~z n In other words, in the case of the Lasso regression method or the Elastic Net regression method, a1 to a n A vector Z' may be selected, which is a set of biomarker candidates of a subset corresponding to terms where the p-value is not 0, and in the case of an AIC / BIC-based regression method, a vector Z'' may be selected, which is a set of candidates selected based on the statistical validity (the p-value, Bonferroni-corrected p-value already mentioned) of the interaction term included in the final regression equation.

[0108] In one example, the sum of the interaction terms included in the final regression equation may be used as a new biomarker candidate. That is, for example, in the case of the above-mentioned [Equation 6], the sum of the interaction terms in the final regression equation is The JPEG2025505384000002.jpg10170 value (numeric value) may be used as a new biomarker.

[0109] 5 is a schematic diagram of a process for generating effect modifier biomarker m and effect modifier-independent biomarker n indicative of the effect of a drug or treatment in a biomarker screening process using an artificial neural network in one embodiment of the present application. Figure 5a shows a case where a numerical vector Z passes through one neural network, and Figure 5b shows a case where the numerical vector Z passes through different neural networks.

[0110] As shown in Figures 5a and 5b, the numerical vector Z may be provided with m and n values ​​through one or more artificial neural networks 31, during which process it may undergo a normalization process 32 and may be passed through a network of transformation functions such as sigmoid, identity, ReLU functions, etc.

[0111] Figure 5a shows that the neural network 31 through which the numerical vector Z passes has one bottom and two heads, where the two heads may have different normalization processes 32 and network structures 31. Figure 5b illustrates that the neural network 31 through which the numerical vector Z passes may be two different neural networks, each having a bottom and a head, which may be trained with separate and independent networks and hyper-parameter sets.

[0112] In one example, various types of artificial neural networks (ANNs) including MLP may be applied as the artificial neural network 31 .

[0113] The biomarkers m and n thus obtained may be used as input values ​​for the following regression equation [Equation 7]. [Formula 7] Q4=a·m+b·m·exposure+c·exposure+d·n+e

[0114] The Q4 is a predicted value for a target (treatment result, e.g., survival or non-survival within a certain period of time, or a value such as weight loss due to treatment) of an exposure (a value related to the quantity and intensity of a medication or treatment, or a value 0 or 1), and the Q4 can be applied to a generalized linear model to train the artificial neural network to minimize the difference (i.e., loss) between the predicted value and the target value, thereby creating an artificial neural network that generates an effect modifier biomarker m. Such biomarker m makes it possible to estimate a treatment effect (corresponding to the b·m·exposure term in Q4) that is influenced by the patient condition, rather than a treatment effect (corresponding to the c·exposure term in Q4) that is unrelated to the patient condition of a particular exposure.

[0115] Here, the loss function for reducing the difference (loss) between the predicted value and the target value may be a cross entropy function or a focal loss function when the exposure is a binary variable, and an MSE function is generally used when the exposure is a numerical variable, but various other loss functions may be used depending on the problem to be solved.

[0116] In one example, the parameters of the regression equation and the neural network are jointly optimized during the training process; for this purpose, various optimization algorithms used in neural network training may be applied, with the selection being based on the final performance of the model.

[0117] On the other hand, effect modifiers can also be generated in the following ways:

[0118] FIG. 6 is a schematic diagram illustrating a process for measuring effect modification in a biomarker screening process in one embodiment of the present application.

[0119] As shown in FIG. 6, when a function h(Z, exposure) that predicts a patient's outcome by taking in a numerical vector Z and, for example, a binary exposure as input, is assumed, two functions may be assumed for the function h when the exposure is 0 (no exposure) or 1 (exposure). That is, g(Z) corresponding to h(Z, 0) when the exposure is 0, and f(Z) corresponding to h(Z, 1) when the exposure is 1 may be generated. Here, when the Z vector is displayed on the X axis (the Z vector may be multidimensional, but is simplified and processed as one dimension and displayed on the X axis) and the outcome is displayed on the Y axis, a function corresponding to each exposure may be shown. Here, the outcome may be, for example, the predicted probability of occurrence of an event (e.g., death), frequency (e.g., frequency of seizures), or a numerical value (e.g., weight change), etc.

[0120] Meanwhile, a separate function m(Z) can be defined to define the treatment effect such as the difference between exposure 1, i.e., f(Z), and exposure 0, i.e., g(Z), with the numerical vector Z on the X-axis and the effect size on the Y-axis (or the ratio between exposure 1 and exposure 0, odds ratio, relative risk, etc., but the calculation method is not limited thereto). Such m(Z) can be calculated and shown as in FIG. 6 (lower figure) to show the change in treatment effect according to the change in vector Z, and can nonlinearly show the effect modifier characteristics of vector Z without generating a separate biomarker. For example, when m(Z)=f(Z)-g(Z), it can be predicted that when numerical vector Z is in a certain range, it will show a strong positive effect (e.g., recovery from disease) for the same treatment, whereas when it is in another specific range, it will show a strong negative effect (e.g., worsening of disease).

[0121] In one embodiment, the f, g, and h functions may be machine learning algorithms, such as, but not limited to, various regression models, neural networks, tree ensembles, SVM, gradient boosting, MLP, etc. Also, known methods for improving the performance of each function may be used.

[0122] Figure 7 is a schematic diagram showing a process of finding attributes of biomarkers in a biomarker screening process in one embodiment of the present application. Figure 7a is a schematic diagram of a characteristic summary of a numeric vector Z in which m biomarker candidates belonging to the numeric vector Z in an m x n matrix each have n attribute vectors X.

[0123] Figure 7b is a diagram illustrating the case where all biomarker candidates are divided into elements of a morphological numerical vector Z', a clinical numerical vector Z'', and a phenotypic numerical vector Z''', and Figure 7c is a diagram illustrating obtaining an attribute vector of a morphological numerical vector based on similarity.

[0124] The feature vector X may be n clinical characteristics (e.g., association with a specific disease, association with prognosis, drug effect, or side effect) of each biomarker as shown in Fig. 7a, or may be measurements of a group of biomarker candidates obtained from n patient cases as shown in Fig. 7b. Such feature vector X for each biomarker candidate element may be stored as an attribute of each biomarker, and may be used to evaluate the similarity between attributes or to cluster biomarkers.

[0125] If a biomarker feature vector X corresponds to measurements obtained from n patient cases, it can be used to generate similarity-based attributes for each of the morphological biomarker candidates. This involves repeatedly calculating the pairwise vector distance / similarity between a vector X created by extracting specific morphological biomarkers belonging to elements of Z' from a set of patient cases, as shown in Figure 7c, and a vector X of interpretable biomarkers (Z' [clinical] or Z'' [phenotypic] biomarkers) extracted from the same cases, thereby extracting multiple similarity scores from one morphological biomarker candidate to obtain a new similarity feature vector X'. This similarity feature vector X' may also be stored as an attribute of the biomarker, and may be used to evaluate the similarity between attributes or to cluster biomarkers.

[0126] If the size (n) of the feature vector is too large in this whole process, the m×n matrix may be subjected to a dimensionality reduction process, for example, using at least one of known dimensionality reduction techniques such as PCA, LDA, SVD, and NMF.

[0127] In one example, the numerical vector Z may be selected as a biomarker through a clustering process. For example, the numerical vector Z may be converted into a K-class categorical variable (where K is the number of clusters) using various clustering techniques such as K-means, K-means++, TDA (topological data analysis), TsNE, and Mixture model, and may be used as a biomarker. For example, the numerical vector Z is extracted from n patients in the same manner to create training data, and an algorithm that defines similar clusters based on the vector Z is trained using the training data, and the algorithm is then applied to a new patient to determine which cluster the patient belongs to. If we know the characteristics (statistics) of each cluster, we can apply the characteristics of the new patient to the new patient to estimate the characteristics of the new patient, and whether or not the patient belongs to a specific cluster can be converted into a K-hot vector and added as storage data for a new prediction model to improve prediction ability.

[0128] Fig. 8 is a schematic diagram illustrating an artificial intelligence-based biomarker selection method in one embodiment of the present application. Fig. 8a is a diagram for explaining each step of the method, and Fig. 8b is a schematic diagram showing a numerical vector, which is a group of biomarker candidates obtained in the process of Fig. 8a, particularly when electrocardiogram data is used, and a method of selecting biomarkers therefrom to obtain biomarkers.

[0129] As shown in FIG. 8a, the artificial intelligence based biomarker selection method may include a step of acquiring biodata (S100), an encoding step (S200) of encoding the biodata to calculate a numerical vector including one or more elements that are biomarker candidates, and a biomarker screening step (S300) of screening biomarkers from one or more elements of the numerical vector.

[0130] The final screened biomarkers obtained from this may be used in various ways, such as prognostic biomarkers, predictive biomarkers, monitoring biomarkers, and surrogate biomarkers.

[0131] Here, a prognostic biomarker is a biomarker that predicts the prognosis of a patient, a predictive biomarker is a biomarker that indicates whether a drug / treatment is effective for the patient, a monitoring biomarker is a biomarker for observing whether a patient will develop problems such as side effects from a drug, and a surrogate biomarker is a biomarker that serves as a substitute for an event that we should observe when it is difficult to directly observe such an event in order to indirectly observe the risk of such an event.

[0132] 8b is a schematic diagram for explaining each step when inputting electrocardiogram data. As described above, a numerical vector that is a biomarker candidate is obtained through encoding (S200), and biomarkers may be selected from the numerical vector through a regression method or the like (S300) to be used as final biomarkers.

[0133] The technology according to the embodiment of the present application described above can be applied to all kinds of diseases other than cardiovascular diseases. In addition, the above-mentioned numerical vector can be used in general artificial intelligence-based medical platforms for diagnosing and predicting various diseases, and can be used as input numerical values ​​that can be utilized in various artificial intelligence algorithms.

[0134] Example Experiment Description 9a to 9c are tables and graphs showing experimental data for evaluating the sensitivity, specificity, etc. of a new electrocardiogram biomarker that predicts the mortality risk of COVID-19-infected patients using 12-lead electrocardiogram (ECG) data and a numerical vector Z extracted from an electrocardiogram analysis encoder in one experimental example of the present invention.

[0135] Referring to Figure 9a, in one experimental example, a total of 167 COVID-19 infected patients admitted to Seoul National University Bundang Hospital were studied. Figure 9a shows basic patient data, and mortality outcome ("Poor") was defined as in-hospital death or application of ECMO.

[0136] A set of 244 candidate biomarkers was extracted from the study subjects' first electrocardiograms taken at their hospitalization using a CNN-based electrocardiogram encoder trained using a combination of self-supervised and supervised learning techniques.

[0137] Figure 9b shows a histogram of the individual predictive performance of the extracted 244 candidate biomarkers.

[0138] As shown in FIG. 9b, the individual prediction performance of the extracted 244 biomarker candidates varied according to the AUC standard, and the highest prediction performance among them was 0.815. In addition, the sensitivity was 87.2%, the specificity was 69.2%, the positive predictive value was 52.6%, and the negative predictive value was 93.3%. Referring to FIG. 9c, the AUC of the best-performing biomarker candidate extracted from the present invention was statistically significantly higher (P<0.05) than the clinical biomarkers SOFA score (0.784) and Troponin I blood concentration (0.770), which are already known to be closely related to patient prognosis in clinical practice. During screening, the screening unit may select the biomarker candidate as a biomarker.

[0139] In addition, the ECG biomarker candidate group and five clinical data including the clinical biomarkers SOFA score and troponin I blood concentration (age, sex, SOFA score, troponin I blood concentration, and Pro-BNP) were nonlinearly combined using a machine learning algorithm (stepwise logistic regression using AIC) to create new biomarker candidates and evaluate their performance. In order to prevent overfitting of the model, 4-fold cross-validation was applied to obtain the final prognostic biomarker. The AUC of the new biomarker obtained in this way was 0.820, which was statistically significantly higher than the AUC of 0.694, which was the performance when the five clinical data were combined without the ECG biomarker (p<0.05).

[0140] The new prognostic biomarker created in this way also showed higher predictive performance than the SOFA score and troponin I blood concentration, which are known to be closely associated with patient prognosis in clinical practice.

[0141] In conclusion, it was found that the group of ECG biomarker candidates extracted using the ECG encoder trained by integrating the self-supervised learning and supervised learning techniques of the present invention through the above experimental examples showed various degrees of predictive ability, and that by combining these with common clinical biomarkers, new biomarkers that show high performance can be created simply by applying a regression model.

[0142] The operations of the methods and devices for implementing the methods according to the embodiments described above may be embodied at least in part in a computer program and recorded in a computer-readable recording medium, for example, with a program product comprising a computer-readable medium containing program code, which may be executed by a processor to perform any or all of the steps, operations, or processes described.

[0143] The computer may be any device that may be or may be integrated into a desktop computer, laptop computer, notebook, smartphone, or similar computing device. A computer is a device that has one or more alternative special purpose processors, memory, storage space, and networking components (either wireless or wired). The computer may run an operating system such as, for example, a Microsoft Windows compatible operating system, Apple OSX or iOS, a Linux distribution, or Google's Android OS.

[0144] The computer-readable recording medium includes all kinds of recording devices in which computer-readable data is stored. Examples of the computer-readable recording medium include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc. The computer-readable recording medium may also be distributed among computer systems connected via a network, and computer-readable code may be stored and executed in a distributed manner. Furthermore, functional programs, codes, and code segments for implementing the present embodiment should be easily understood by those skilled in the art to which the present embodiment pertains.

[0145] The present invention has been described above with reference to the embodiments shown in the drawings, but these are merely illustrative, and those skilled in the art will understand that various modifications and variations of the embodiments are possible based on the above. However, such modifications should be considered to be within the technical scope of the present invention. Therefore, the true technical scope of the present invention should be determined by the technical ideas of the appended claims. [Industrial Applicability]

[0146] The present application relates to an artificial intelligence-based biomarker selection device and method capable of developing companion diagnostic biomarkers that can be utilized in new drug development, precision medicine, disease diagnosis, and treatment optimization for various diseases, including those for which no existing biomarkers exist, such as cardiovascular diseases.

[0147] This application can be utilized for the discovery of various biomarkers that can be used for drug discovery, such as the development of new cardiovascular drugs, prognosis prediction, etc. This can be effectively utilized in pharmaceutical platforms for new drug development, and research platforms for precision medicine, disease diagnosis, treatment optimization, etc.

Claims

1. 1. An artificial intelligence based biomarker selection device, comprising: an acquisition unit for acquiring biodata; An encoder that receives the biodata and calculates a numerical vector including one or more elements that are biomarker candidates; and An artificial intelligence-based biomarker selection device, comprising: a biomarker screening unit that screens biomarkers from one or more elements of the numerical vector.

2. The artificial intelligence-based biomarker selection device according to claim 1 , wherein the biodata is two or more biodata of the same or different modalities.

3. The artificial intelligence-based biomarker selection device of claim 2, wherein the biodata includes at least one of signal data and image data.

4. 4. The artificial intelligence based biomarker selection device of claim 3, wherein the encoder is trained in a manner that projects each of the numerical vectors calculated by position embedding a plurality of signals and / or images onto a common space using another artificial neural network that is trained together, and finds positive pairs - pairs of the same subject within a specified time range; and / or negative pairs - pairs of the same subject outside a specified time range or pairs between different subjects - based on the similarity between the numerical vectors.

5. The artificial intelligence-based biomarker selection device of claim 4, wherein the positive pair or the negative pair includes one or more of a signal-signal pair, an image-image pair, and a signal-image pair.

6. The artificial intelligence-based biomarker selection device of claim 4, wherein the encoder includes an NCE loss as a loss function.

7. The artificial intelligence-based biomarker selection device according to claim 6, wherein the loss function further includes at least one of a MIL NCE loss and a supervised task loss.

8. 8. The artificial intelligence-based biomarker selection device of claim 7, wherein the supervised task is a task related to prediction, and is a task of simultaneously or sequentially predicting one or more of binary, multi-category, and numerical values.

9. the encoder is trained with a plurality of signals and / or images; The artificial intelligence-based biomarker selection device of claim 4, wherein the plurality of signals and / or images are subjected to data augmentation processing by applying original signal and / or imaging transformation.

10. The encoder includes a first encoder receiving an input of a data-augmented first biodata set; and a second encoder receiving an input of a data-augmented second biodata set; the first biodataset and the second biodataset are of the same modality; The artificial intelligence-based biomarker selection device of claim 9, wherein the second encoder shares the same weight as the first encoder or is a momentum encoder that averages the temporal changes of the first encoder.

11. The first encoder may or may not include an MLP layer; The artificial intelligence-based biomarker selection device of claim 10, wherein the second encoder performs or does not perform clustering to replace the calculated numerical vector with an embedded numerical vector representative value.

12. At least one of the first encoder and the second encoder calculates a similarity loss between positive pairs according to a similarity function; and / or The artificial intelligence-based biomarker selection device of claim 11, wherein at least one of the first encoder and the second encoder calculates a dissimilarity loss between negative pairs by a dissimilarity function.

13. 2. The artificial intelligence-based biomarker selection device of claim 1, wherein the encoder comprises at least one of a clinical encoder and a morphological encoder, the clinical encoder being trained by supervised learning, and the morphological encoder being trained by unsupervised learning.

14. The artificial intelligence-based biomarker selection device of claim 13, wherein the encoder uses a numerical vector obtained by concatenating a numerical vector calculated from a clinical encoder and a numerical vector calculated from a morphological encoder, and is trained by multi-task learning in which supervised learning and unsupervised learning are integrated.

15. The artificial intelligence-based biomarker selection device of claim 1, characterized in that at least one or more of the mean, range, and standard deviation (SD) of some or all of one or more elements of each numerical vector repeatedly calculated from the encoder are included in the biomarker candidates.

16. The artificial intelligence-based biomarker selection device according to claim 1 , wherein the biomarker screening unit stores one or more elements of the numerical vector that are the biomarker candidates together with attribute information.

17. The artificial intelligence-based biomarker selection device of claim 1, wherein at least a portion of one or more elements of the numerical vector are combined through linear or nonlinear transformation and used as biomarker candidates.

18. The artificial intelligence-based biomarker selection device of claim 17, wherein at least a portion of one or more elements of the numerical vector are combined with a clinical biomarker via linear or nonlinear transformation and used as a biomarker candidate.

19. The artificial intelligence-based biomarker selection device of claim 1, wherein a numerical value obtained through a supervised task related to the prediction of the encoder is concatenated into the numerical vector.

20. The artificial intelligence-based biomarker selection device of claim 1 , wherein the numerical vector includes a morphological numerical vector, a clinical numerical vector, and basic patient data.

21. 2. The artificial intelligence based biomarker selection device according to claim 1, wherein the biomarker screening unit screens biovectors from biomarker candidates through one or more of regression, decision tree, clustering, dimension reduction, and supervised learning.

22. The artificial intelligence-based biomarker selection device according to claim 20, wherein the biomarker screening unit selects one or more elements that are biomarker candidates as a group by grouping them as a biomarker.

23. The biomarker screening unit comprises: The artificial intelligence-based biomarker selection device according to claim 22, characterized in that one or more elements of a numerical vector related to a learning task related to prediction are selected as biomarkers, and / or one or more elements of a numerical vector unrelated to a learning task related to prediction are selected as biomarkers using a clustering technique based on a similarity between the elements and a predictor element related to the learning task related to prediction.

24. The biomarker screening unit selects a numerical vector, which is a biomarker candidate, as a biomarker using the following regression equation [Equation 4], [Formula 4] Q1=af(Z)+c In the above formula, f is a linear or nonlinear transformation function including an identity function without any specific processing, the absolute value of the coefficient a indicates the effect size, and c is an intercept; Evaluate whether the numerical vector Z is statistically significant with respect to the result Q1 using the regression equation of [Equation 4]; The artificial intelligence-based biomarker selection device according to claim 22, characterized in that, when the absolute value of the coefficient a, which is an effect size, is equal to or greater than a preset value and the p-value of the coefficient a is less than a preset value, the numerical vector Z is selected as a biomarker.

25. The artificial intelligence based biomarker selection device of claim 24, wherein the biomarker screening unit further selects, as biomarkers, numerical vectors whose coefficient values ​​are not equal to 0 through a Lasso regression method and / or an Elastic Net regression method for the selected numerical vectors.

26. The biomarker screening unit selects biomarkers using the following regression equation [Equation 5], [Formula 5] Q2 = a z ｎ ・Exposure + b・z ｎ +c Exposure +d where a, b, and c are coefficients, d is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; In the regression equation of the above [Formula 5], the biomarker candidate group z １ ..z ｎ assess whether the effect modification of the exposure on the target outcome is statistically significant; When the absolute value of the coefficient a, which is the effect size, is equal to or greater than a preset value and the p-value of the coefficient a is less than a preset value, the biomarker z ｎ The artificial intelligence-based biomarker selection device according to claim 22, characterized in that it selects as a biomarker.

27. The biomarker screening unit is configured to screen one or more candidate biomarkers z １ , z ２ ..z ｎ A biomarker is selected using the following regression equation [Equation 6] for [Formula 6] Q3 = Exposure (a １・ f １ (z １ ) + a ２・ f ２ (z ２ ) +...+a ｎ・ f ｎ (z ｎ )) + b １・ g １ (z １ ) + b ２・ g ２ (z ２ ) +...+b ｎ・ g ｎ (z ｎ ) + c ・ Exposure +d In the above formula, the f function and the g function are linear or nonlinear transformation functions including an identity function, １ , a ２ ..a ｎ , b １ , b ２ ..b ｎ , and c are coefficients, d is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; When the Lasso regression method or the Elastic Net regression method is used as the regression method, １ ~a ｎ A new Z′ vector composed of a set of biomarker candidates having a coefficient that is not zero until is selected as a biomarker; The artificial intelligence-based biomarker selection device according to claim 26, characterized in that when AIC- and BIC-based regression methods are used as regression methods, a new Z'' vector composed of a set of biomarker candidates selected based on the statistical validity of interaction terms included in the final regression equation is selected as a biomarker.

28. The sum of the interaction terms in the regression equation of [Equation 6] 28. The artificial intelligence-based biomarker selection device of claim 27, further using the value (numeric value) as a new biomarker.

29. The biomarker screening unit generates an effect modifier using the following regression equation [Equation 7]: [Formula 7] Q4 = a.m + b.m.exposure + c.exposure + d.n + e where a, b, c, d are coefficients, e is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; m is an effect modifier biomarker that indicates an effect for a particular drug or treatment selected by passing the numerical vector Z through an artificial neural network, and n is a biomarker that is unrelated to effect modification; 23. The device for selecting a biomarker based on artificial intelligence according to claim 22, wherein the predicted value Q4 obtained by the regression equation of [Equation 7] is applied to a generalized linear model to train the artificial neural network so as to minimize a difference between the predicted value and a target value, thereby generating an m value, which is an effect modifier biomarker, and selecting the m value as a biomarker.

30. The difference between the predicted value and the target value is If the exposure is a binary variable, a cross entropy loss function or a focal loss function is used, 30. The artificial intelligence-based biomarker selection device of claim 29, wherein an MSE function is used when the exposure is a numerical variable.

31. 23. The artificial intelligence based biomarker selection device of claim 22, wherein the biomarker screening unit selects biomarkers from at least one of a difference between two function values, a ratio between two function values, an odds ratio, and a relative risk when the exposure is 0 or 1 in a function h(Z, exposure) that receives a numerical vector Z and a binary exposure as input and predicts a patient's treatment outcome.

32. The biomarker screening unit may form an m×n matrix composed of m biomarker candidates and n feature vectors X, and the feature vector X may be a feature possessed by each of the m biomarker candidates; The artificial intelligence-based biomarker selection device of claim 16, wherein one or more elements of feature vector X present for each of m biomarker candidates in the m×n matrix are stored as attributes of the biomarkers and are used to evaluate similarity between attributes or to cluster the biomarkers.

33. The artificial intelligence-based biomarker selection device according to claim 32, wherein the matrix is ​​used after dimension reduction.

34. The artificial intelligence-based biomarker selection device of claim 32, characterized in that the feature vector X present for each of the m biomarker candidates includes a morphological numerical vector, a clinical numerical vector, and a phenotype vector, and a vector constructed using the similarity between the morphological numerical vector and the clinical numerical vector and / or the similarity between the morphological numerical vector and the phenotype vector as elements is used as an attribute of the morphological numerical vector.

35. The artificial intelligence-based biomarker selection device of claim 1, wherein the encoder uses at least one of a Bayesian layer and a KL loss function to reduce correlation between multiple elements.

36. The biomarker is predictive biomarkers that can distinguish between responders and non-responders to a particular drug; and / or The artificial intelligence-based biomarker selection device according to claim 1, characterized in that the biomarker is a prognostic biomarker capable of predicting the prognosis of a disease.

37. The artificial intelligence-based biomarker selection device according to any one of claims 1 to 36, characterized in that the biodata is electrocardiogram (ECG) data.

38. 38. The artificial intelligence-based biomarker selection device of claim 37, wherein the biomarker is a biomarker related to cardiovascular disease.

39. 1. An artificial intelligence based biomarker selection method, comprising: An encoding step of encoding the biodata to generate a numerical vector including one or more elements that are biomarker candidates; and A method for selecting biomarkers based on artificial intelligence, comprising: a biomarker screening step of screening biomarkers from one or more elements of the numerical vector.

40. The encoding step includes projecting the respective numerical vectors calculated by position embedding a plurality of signals and / or images onto a common space using another artificial neural network trained together, and training the artificial neural network in a manner that finds positive pairs - pairs of the same subject within a specified time range; and / or negative pairs - pairs of the same subject outside a specified time range or pairs between different subjects - based on the similarity with the numerical vectors. The artificial intelligence-based biomarker selection method according to claim 39, characterized in that the loss function during training includes an NCE loss, and further includes at least one of an MIL NCE loss and a supervised task loss.

41. the encoding step is performed using different encoders trained using a plurality of signals or images of the same modality; The different encoders share the same weight or use a momentum encoder that averages the time variation of one encoder as the other encoder; The encoding step may or may not include an MLP layer; The encoding step is to perform clustering by replacing the calculated numeric vector with an embedded numeric vector representative value, or not to perform clustering, 40. The method of claim 39, wherein the encoding step includes calculating a similarity loss between positive pairs using a similarity function and / or calculating a dissimilarity loss between negative pairs using a dissimilarity function.

42. 40. The artificial intelligence based biomarker selection method according to claim 39, wherein the biomarker screening step comprises screening biovectors from biomarker candidates through one or more of regression, decision tree, clustering, dimension reduction, and supervised learning.

43. The biomarker screening step comprises: Selecting one or more elements of a numerical vector related to a supervised task related to the prediction as a biomarker, and / or The artificial intelligence-based biomarker selection method according to claim 42, characterized in that one or more elements of a numerical vector unrelated to a learning task related to a prediction are selected as biomarkers using a clustering technique based on their similarity to a predictor element related to the learning task related to the prediction.

44. The biomarker screening step is to select a biomarker candidate numerical vector as a biomarker using the following regression equation [Equation 4]: [Formula 4] Q1=af(Z)+c In the above formula, f is a linear or nonlinear transformation function including an identity function without any specific processing, the absolute value of the coefficient a indicates the effect size, and c is an intercept; Evaluate whether the numerical vector Z is statistically significant with respect to the result Q1 using the regression equation of [Equation 4]; The artificial intelligence-based biomarker selection method according to claim 42, characterized in that if the absolute value of the effect size coefficient a is equal to or greater than a preset value and the p-value of the effect size coefficient a is less than a preset value, the numerical vector Z is selected as a biomarker.

45. The biomarker screening step is to select a biomarker using the following regression equation [Equation 5]: [Formula 5] Q2 = a z ｎ ・Exposure + b・z ｎ +c Exposure +d where a, b, and c are coefficients, d is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; Evaluate whether the effect modification of the exposure on the target outcome is statistically significant for the numerical vector Z, which is a group of biomarker candidates, in the regression equation of [Equation 5]; When the absolute value of the coefficient a, which is the effect size, is equal to or greater than a preset value and the p-value of the coefficient a is less than a preset value, the biomarker z ｎ The artificial intelligence-based biomarker selection method of claim 42, characterized in that the following is selected as a biomarker:

46. The biomarker screening step comprises screening one or more candidate biomarkers z １ , z ２ ..z ｎ A biomarker is selected using the following regression equation [Equation 6] for [Formula 6] Q3 = Exposure (a １・ f １ (z １ ) + a ２・ f ２ (z ２ ) +...+a ｎ・ f ｎ (z ｎ )) + b １・ g １ (z １ ) + b ２・ g ２ (z ２ ) +...+b ｎ・ g ｎ (z ｎ ) + c ・ Exposure +d In the above formula, the f function and the g function are linear or nonlinear transformation functions including an identity function, １ , a ２ ..a ｎ , b １ , b ２ ..b ｎ , and c are coefficients, d is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; When the Lasso regression method or the Elastic Net regression method is used as the regression method, １ ~a ｎ A new Z′ vector composed of a set of biomarker candidates having a coefficient that is not zero until is selected as a biomarker; When using AIC and BIC-based regression methods as regression methods, a new Z'' vector composed of a set of biomarker candidates selected based on the statistical validity of the interaction term included in the final regression equation is selected as a biomarker; The sum of the interaction terms in the regression equation of [Equation 6] 43. The artificial intelligence-based biomarker selection method of claim 42, further comprising using the value (numeric value) as a new biomarker.

47. The biomarker screening step is to generate an effect modifier using the following regression equation [Equation 7]: [Formula 7] Q4 = a.m + b.m.exposure + c.exposure + d.n + e where a, b, c, d are coefficients, e is the intercept, and exposure is a binary or other numeric variable value relating to exposure to a particular drug or treatment; m is an effect modifier biomarker that indicates an effect for a particular drug or treatment selected by passing the numerical vector Z through an artificial neural network, and n is a biomarker that is unrelated to effect modification; The predicted value Q4 obtained by the regression equation of [Equation 7] is applied to a generalized linear model to train the artificial neural network so as to minimize the difference between the predicted value and the target value, thereby generating an m value, which is an effect modifier biomarker, and selecting it as a biomarker; The difference between the predicted value and the target value is If the exposure is a binary variable, a cross entropy loss function or a focal loss function is used, 43. The artificial intelligence-based biomarker selection method of claim 42, wherein an MSE function is used when the exposure is a numerical variable.

48. 43. The method of claim 42, wherein the biomarker screening step selects biomarkers from at least one of a difference between two function values, a ratio between two function values, an odds ratio, and a relative risk when the exposure is 0 or 1 in a function h(Z, exposure) that predicts a patient's treatment outcome by inputting a numerical vector Z and a binary exposure.

49. The biomarker screening step may include forming an m×n matrix composed of m biomarker candidates and n feature vectors X, the feature vector X being a feature possessed by each of the m biomarker candidates; In the m×n matrix, one or more elements of a feature vector X having n elements present for each of m biomarker candidates are stored as attributes of the biomarkers and are used to evaluate the similarity between attributes or to cluster the biomarkers. The artificial intelligence-based biomarker selection method according to claim 39, wherein the matrix is ​​used after dimension reduction.

50. The method for selecting a biomarker based on artificial intelligence according to claim 49, wherein the n feature vectors X present for each of the m biomarker candidates include a morphological numerical vector, a clinical numerical vector, and a phenotype vector, and a vector constructed using the similarity between the morphological numerical vector and the clinical numerical vector and / or the similarity between the morphological numerical vector and the phenotype vector as elements is used as an attribute of the morphological numerical vector.

51. The artificial intelligence-based biomarker selection method of claim 39, wherein the encoding step uses at least one of a Bayesian layer and a KL loss function to reduce correlation between multiple elements.

52. A computer program stored on a medium for performing, in combination with hardware, the artificial intelligence-based biomarker selection method according to any one of claims 39 to 51.

Images