KR102182807B1 - Apparatus of mixed effect composite recurrent neural network and gaussian process and its operation method - Google Patents

Abstract

A complex model device using a recurrent neural network and a Gaussian process and an operation method thereof are presented. According to an embodiment, a complex model apparatus using a recurrent neural network and a Gaussian process includes: a recurrent neural network (RNN) that obtains a shared characteristic, which is an overall trend among a plurality of heterogeneous patients, from time series data; And a Gaussian Processes (GP) modeling individual characteristics for each patient from the time series data, and the time series data can be personalized and reliably predicted through the recurrent neural network and the Gaussian process.

Description

Complex model device using recurrent neural network and Gaussian process and its operation method {APPARATUS OF MIXED EFFECT COMPOSITE RECURRENT NEURAL NETWORK AND GAUSSIAN PROCESS AND ITS OPERATION METHOD}

The following embodiments relate to a personalized and reliable healthcare prediction model in the medical field, and more particularly, to a complex model apparatus using a recurrent neural network and a Gaussian process, and an operation method thereof.

Precision medicine provides personalized medical services such as diagnosis, disease treatment and prevention, and the ultimate goal of healthcare is precision medicine. While difficult in the past, these days, data-driven approaches such as machine learning have evolved and are becoming increasingly feasible. In particular, electronic health records (EHR), which is a systematic collection of patients' various clinical records, have been widely used in recent years, which machine learning researchers have used to determine the quality of clinical treatment based on records of personal medical history. In order to improve it, various clinical reasoning models are being studied.

Some machine learning applications using electronic health records (EHR) are already having success in a number of clinical tasks, such as predicting heart failure risk, predicting sepsis, and analyzing physiological time series. In particular, with the recent success of deep learning, recurrent neural networks (RNNs) are becoming the most popular option when studying electronic health record (EHR) data.

However, since it is difficult to grasp the contribution and uncertainty of each input to the prediction of each input, the recurrent neural network (RNN) generally lacks image analysis capability, which is fatal to the clinical task. In addition, since deep networks generally require enormous data due to large-scale parameters, it is very difficult to train a separate model for each patient. Therefore, if your goal is to predict the mean well, a cyclic neural network (RNN) may be a reasonable option, but the prediction will not be optimized for each patient.

), 성능 감소를 수반하며 이러한 문제를 완화하기 위해 몇몇 근사 방법이 제안되었지만 정확한 가우시안 프로세스(GP)를 모델링하는 것은 계산적으로 매우 어렵다.The Gaussian Process (GP) is another popular model for time series. Due to its probabilistic nature, the Gaussian process (GP) is sometimes used to indicate that uncertainty is important. However, when doing research using large-scale data, it is necessary to calculate the inverse of the covariance matrix over all data points (for accurate inference for n data points

), performance reduction, and some approximation methods have been proposed to alleviate this problem, but it is computationally very difficult to model an accurate Gaussian process (GP).

A significant limitation of existing approaches (e.g., a method called "one-model-fits-all") is that when forecasting ignores the large variability of patients among unobserved patient properties, only observed factors are considered. Point. This diversity arises from a variety of causes, such as intrinsic differences due to demographic and biological factors or other environmental differences. When performing a diagnosis, such intrinsic differences between patients may not imply that two equivalent clinical features are equivalent to the progression of the target disease. Also, randomly divided data points for each patient can be a problem when applied to a single model such as a recurrent neural network (RNN). Nonetheless, these methods treat patient collections in an electronic health record (EHR) only as independent and identically distributed observations and provide a single in-depth model (or personalized Gaussian sharing parameters to reduce computational cost). Process (GP) model).

In order to explain this problem, it is explained through FIG. 1 how heterogeneities between patients can affect the overall estimation. 1 is a diagram for explaining the influence of heterogeneity among patients according to an exemplary embodiment.

를 통해 환자

에 대한 데이터를 생성할 수 있다. 이 함수는 모든 환자들이 공유하고 있는 전반적인 경향(global trend) 뿐만 아니라, 다른 이들과의 편차를 만드는 환자 맞춤(patient-specific) 노이즈

를 포함할 수 있다. 여기서 학습 데이터가 어떤 제한된 기간(x축 5 부분에 도시된 점선 이전)동안 수집된다고 가정할 수 있다. 여기에서의 목적은 보이는 입력 범위(점선 이전)뿐만 아니라, 보이지 않는 입력 범위(점선 이후)에서도 새로운 환자들에 대한 출력을 예측하는 것이다. As shown in Figure 1, in the simulation

Patient through

You can create data for This function not only provides a global trend shared by all patients, but also patient-specific noise that creates deviations from others.

It may include. Here, it can be assumed that the training data is collected for a certain limited period (before the dotted line shown in the x-axis 5). The purpose here is to predict the output for new patients over the visible input range (before the dotted line) as well as the invisible input range (after the dotted line).

다양성에 기인하여 모든 최신 기술들의 새로운 환자에 대한 전반적인 추정이 급격하게 저하되는 것을 볼 수 있다. 이는 전반적인 경향뿐만 아니라 개인적인 특성을 무시하는 것의 위험성을 시사한다.This shows that all models fit well with the patients included in the training set. But the patients

Due to the versatility, it can be seen that the overall estimates of new patients in all state-of-the-art technologies degrade dramatically. This suggests the dangers of ignoring personal characteristics as well as overall tendencies.

Choi, E.; Bahadori, M. T.; Stewart, W.; and Sun, J. 2016a. Doctor AI: Predicting clinical events via recurrent neural networks. In Machine Learning for Healthcare. Nickisch, H., and Rasmussen, C. E. 2008. Approximations for binary Gaussian process classification. JMLR 9.

The embodiments describe a complex model apparatus using a recurrent neural network and a Gaussian process, and an operation method thereof, and more specifically, provide a technique capable of providing uncertainty about prediction by combining a recurrent neural network with a Gaussian process model.

In addition, the embodiments provide a framework for creating a personalized model for each patient, thereby efficiently modeling the patient characteristics hidden in the data and providing a personalized prediction for each patient's characteristics, a complex model using a recurrent neural network and a Gaussian process. It is to provide an apparatus and a method of operation thereof.

According to an embodiment, a complex model apparatus using a recurrent neural network and a Gaussian process includes: a recurrent neural network (RNN) that obtains a shared characteristic, which is an overall trend among a plurality of heterogeneous patients, from time series data; And a Gaussian Processes (GP) modeling individual characteristics for each patient from the time series data, and the time series data can be personalized and reliably predicted through the recurrent neural network and the Gaussian process.

It may further include an optimizer for minimizing the sum of the set of negative marginal log-likelihood (negative marginal log-likelihood) of the Gaussian process obtained from the patients.

The Gaussian process may probabilistically model the individual characteristics when a relatively small number of time points are given, and may provide uncertainty for prediction.

The optimizer may minimize the sum of the set of negative marginal log-likelihoods by using the shared characteristics obtained from the recurrent neural network and the individual characteristics obtained from the Gaussian process.

The time series data may be an electronic health record (Electronic Health Care, EHR).

A single Gaussian process model is constructed for each patient by acquiring shared characteristics through the recurrent neural network and reflecting individual characteristics through the Gaussian process, and the single Gaussian process model is a personalized prediction for each patient's characteristics. Can provide.

According to another embodiment, a method of operating a complex model apparatus using a recurrent neural network and a Gaussian process includes: obtaining a shared characteristic, which is an overall trend among heterogeneous patients, from time series data using a recurrent neural network (RNN); And modeling individual characteristics for each patient from the time series data using Gaussian Processes (GP), and the time series data may be personalized and reliably predicted through the recurrent neural network and the Gaussian process. .

According to embodiments, a complex model device using a recurrent neural network and a Gaussian process capable of providing uncertainty about prediction by combining a Gaussian process model with a recurrent neural network for personalized and reliable healthcare prediction, and an operation method thereof. I can.

In addition, according to embodiments, by providing a framework for creating a personalized model for each patient, the characteristics of patients hidden in the data are efficiently modeled, and a recurrent neural network and a Gaussian process are used to provide personalized predictions for each patient's characteristics. It is possible to provide a composite model device and an operation method thereof.

1 is a diagram for explaining the influence of heterogeneity among patients according to an exemplary embodiment.
FIG. 2 is a diagram illustrating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.
3 is a flowchart illustrating a method of operating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.
4 is a block diagram illustrating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.
5 shows a graphic representation of a multi-task effect framework according to an embodiment.
6 shows a graphical representation of a composite model (MECGP) using a recurrent neural network and a Gaussian process according to an embodiment.
7 shows a graphical representation of an individual Gaussian process (GP) model for a patient according to one embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in various forms, and the scope of the present invention is not limited by the embodiments described below. In addition, various embodiments are provided to more completely explain the present invention to those of ordinary skill in the art. In the drawings, the shapes and sizes of elements may be exaggerated for clearer explanation.

The following examples present a personalized and reliable healthcare predictive model capable of providing individual personalized medical services such as diagnosis, disease treatment and prevention. The proposed framework aims to make personalized and reliable predictions by modeling two complementary elements from time-series data such as Electronic Health Care (EHR).

To this end, a deep recurrent neural network (RNN) to use the expressive power of a recurrent neural network (RNN) to estimate overall trends from many patients, and a relatively small number of time points (time points). point), we propose a complex model for Gaussian Processes (GP) to probabilistically model individual time series.

FIG. 2 is a diagram illustrating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.

First, looking at the properties of medical data, such as electronic health records (EHR), medical data is very noisy (heterogeneous). This is mainly due to the huge gap between the demographic and biological hidden factors across the patient. In addition, medical data is collected in a time series manner, and at this time, the patient's past data is useful to the doctor. In addition, treatment (results) through medical data must be reliable.

As can be seen from the nature of medical data, a single model may not be suitable for various (heterogeneous) patients, and in heterogeneous medical situations, individualized multi-task models may be required for each patient. And, medical treatment must be reliable, and deep learning methods generally work well with medical data. However, models that cannot provide predictive uncertainty may not be clinically practical.

Referring to FIG. 2, embodiments propose a complex model apparatus 100 using a recurrent neural network (RNN) and a Gaussian process (GP) having the advantages of both models, and this is a complex model using a recurrent neural network and a Gaussian process ( It can be called Mixed-Effect Composite RNN-Gaussian Processes, MECGP (100). Using the complementary characteristics of the two models (the recurrent neural network model and the Gaussian process model) in a composite model (MECGP) using a recurrent neural network and a Gaussian process (100), obtain a global trend among numerous heterogeneous patients. We can use a recurrent neural network (RNN) for this and a Gaussian process (GP) to model each patient-specific variability.

In particular, the composite model (MECGP) 100 using a recurrent neural network and a Gaussian process is a multi-task framework that provides personalized and reliable inference for time series data such as electronic health records (EHR). The (or patient-level) process models two factors: i) the one that acquires the shared trait, which is the global trend of various patients, and ii) the individual trait, which is the unique variability for each patient. It is composed of the sum of the elements.

Since targeting time-series medical datasets such as Electronic Health Records (EHR) dealing with countless different types of patients, this sharing process must be mathematically effective to learn the overall trends of patients. In addition, it should be able to handle countless patients easily computationally using computers during learning as well as reasoning time. Here, a recurrent neural network (RNN) is used to utilize the mathematical ability and relative computational efficiency. On the other hand, for the latter, each process must be maintained separately for each patient without any information sharing, so it must be reliable even if there are a very limited number of data points (where each time point in the data points to a patient's hospital visit). ). Thus, we select the Gaussian process (GP) as an individual model for each patient. A Gaussian process (GP) can be used to naturally obtain prediction uncertainty. This prediction uncertainty is another important feature for reliable predictions for mission-critical clinical tasks.

For 12 common diseases obtained from a large electronic health record (EHR) dataset, a classification and regression task including a risk prediction task is performed to evaluate the proposed model 100. The comparative analysis proves that the composite model (MECGP) 100 using a recurrent neural network and a Gaussian process is significantly better than current state-of-the-art methods including deep learning approaches, and is able to express predictive uncertainty for each patient.

The complex model using a recurrent neural network and a Gaussian process according to an embodiment begins with the following three research areas: variants of a Gaussian process model, a deep model, and a combination of a deep model and a Gaussian process (GP). The following focuses on studies related to modeling electronic health records (EHR) rather than describing the comprehensive study.

First, multiple Gaussian Processes using Electronic Health Record (EHR) will be described.

Gaussian Process models have been actively used in medical applications because of their reliability and versatility. However, it is preferred to use a separate formulation for multiple Gaussian processes (GPs) due to computational cost concerns. In order to solve the missing values due to sensor artifacts or data incompleteness, which are common situations in wearable devices, multiple Gaussian process formulations have been proposed. In addition, it has been proposed to use a similar model for the diagnosis of Alzheimer, which applies a population-level Gaussian process (GP) to new patients individually using the domain Gaussian process (GP). This model can be understood as multi-task learning in the sense that the parameters of the patients Gaussian process (GP) are shared. Most studies using multiple Gaussian process (GP) as multi-task learning utilize Gaussian process (GP) in a separate way or use Gaussian process (GP) to share parameters, which is a common knowledge of tasks. Is not fully available.

Subsequently, a multi-task Gaussian process using electronic health record (EHR) will be described.

Several models have proposed a multi-task Gaussian process with covariance functions shared among take-specific models. Conventionally, for multi-task learning, a multi-task Gaussian process (MTGP), which uses an input-specific covariance matrix and a task-specific covariance matrix is combined. Suggested. A practical example of applying MTGP to a medical situation is to correlate multivariate physiological time series data. In addition, in the past, the covariance matrix is structured with the Linear Model of Coregionalization (LMC) framework, and another approach has been proposed to share the covariance function shared between personalized Gaussian processes (GPs) for individual patients. Previously, MTGP was used to pre-process the input data entering the recurrent neural network (RNN). All these studies are multi-task Gaussian processes (GPs) based on shared covariance functions, which make the total covariance matrix too large, making the inference difficult to handle in an accurate way.

However, the model proposed according to this embodiment is built with a differentiated method of acquiring a complex structure using a flexible deep model for shared mean functions, and more importantly, each patient is used to reflect individual signals. Is that it clearly establishes a single Gaussian process (GP) for.

Next, a deep learning model Recurrent Neural Networks (RNN) using electronic health records (EHR) will be described.

Recently, it has been in the spotlight as a means to train predictive models using time-series clinical data such as electronic health records (EHR). (Non-Patent Document 1) is a Long-Short Term Memory for multi-label classification of diagnostic codes when multivariate features are given from an electronic health record (EHR). , LSTM) and Gated Recurrent Units (GRU). In addition, in order to further improve the performance of models by introducing the concept of missing indicator and decaying informativeness, the pattern of missingness, a typical property of electronic health records (EHR), was used. . In addition, it was previously proposed to build an interpretable model and to use a recurrent neural network (RNN) to generate a degree of attention on which features and hospital visits the model should handle, and to prove its performance using a heart failure prediction task. did. Although the recurrent neural network (RNN) model has shown impressive performance on real-world clinical datasets, the use of a recurrent neural network (RNN) model for essential safety clinical tasks must be done with caution because of the lack of certainty or uncertainty of predictions.

In addition, a purely deterministic recurrent neural network (RNN), which combines the deep model and the Gaussian process (GP), will be described.

Since it is not possible to give confidence in predictions, there has been a lot of interest recently in inducing prediction uncertainty using techniques in the Bayesian statistics field. One of the direct efforts to achieve this goal in the healthcare field is to combine deep architectures and the Gaussian Process to benefit from the benefits of both models. Previously, a Gaussian process (GP) was placed at the output of a deep network to obtain more expressive power, and a concept similar to that of modifying the covariance function of the Gaussian process (GP) was studied. To achieve the same goal of obtaining a more expressive model, deep networks were formulated as stacked Gaussian Processes.

The embodiments aim at combining a deep neural network and a Gaussian process (GP), but the proposed composite model is a more efficient method of using the strengths of two complementary models. In particular, embodiments utilize a deep network to obtain a global complex structure in data, and a Gaussian process (GP) may be used to obtain local variability within an individual instance.

3 is a flowchart illustrating a method of operating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.

Referring to FIG. 3, a method of operating a complex model device using a recurrent neural network and a Gaussian process according to an embodiment is a shared characteristic, which is an overall trend between patients who are heterogeneous from time series data using a recurrent neural network (RNN). Obtaining (S110), and modeling the individual characteristics for each patient from the time series data using a Gaussian process (Gaussian Processes, GP) (S120), and the time series data through a recurrent neural network and a Gaussian process. Personalized and reliably predictable.

In addition, an optimization step (S130) of minimizing the sum of a set of negative marginal log-likelihoods of Gaussian processes obtained from patients may be further included.

Embodiments may provide uncertainty about prediction by combining a recurrent neural network with a Gaussian process model. In addition, the embodiments provide a framework for creating a personalized model for each patient, thereby efficiently modeling the characteristics of patients hidden in the data and providing personalized predictions for the characteristics of each patient.

In the following, each step of a method of operating a complex model device using a recurrent neural network and a Gaussian process will be described in more detail.

A method of operating a complex model device using a recurrent neural network and a Gaussian process according to an embodiment may be described in more detail with one example through the complex model device using a recurrent neural network and a Gaussian process according to an embodiment.

4 is a block diagram illustrating a complex model apparatus using a recurrent neural network and a Gaussian process according to an embodiment.

Referring to FIG. 4, a composite model apparatus 100 using a recurrent neural network and a Gaussian process according to an embodiment may include a recurrent neural network 110, a Gaussian process 120, and an optimizer 130.

In step S110, the recurrent neural network 110 may acquire a shared characteristic, which is an overall trend among a plurality of heterogeneous patients, from the time series data. Here, the time series data may be an electronic health record.

In step S120, the Gaussian process 120 may model individual characteristics for each patient from time series data. The Gaussian process 120 may probabilistically model individual characteristics when a relatively small number of time points are given, and may provide uncertainty for prediction.

In step S130, the optimizer 130 may minimize the sum of a set of negative marginal log-likelihoods of the Gaussian process 120 obtained from patients. The optimizer 130 minimizes the sum of the set of negative marginal log-likelihoods by using the shared characteristics obtained from the recurrent neural network 110 and the individual characteristics obtained from the Gaussian process 120. I can.

As described above, time series data can be personalized and reliably predicted through the recurrent neural network 110 and the Gaussian process 120.

That is, a single Gaussian process model can be built for each patient by acquiring the shared characteristics through the recurrent neural network 110 and reflecting the individual characteristics through the Gaussian process 120, and such a single Gaussian process model It can provide personalized predictions for star characteristics.

Hereinafter, the complex model apparatus 100 using a recurrent neural network and a Gaussian process and an operation method thereof will be described in more detail. Here, the complex model apparatus 100 using a recurrent neural network and a Gaussian process may be referred to as a framework of multi-task mixed effect models.

명의 환자로 구성되며,

번째 환자는

원소들(또는 방문)의 시퀀스로 표현된다고 가정할 수 있다. 즉,

이며, 대응하는 목표 값

이다. 환자 모델링 시, 예측 태스크의 목표는 이전의 모든 입력 특징들 및 목표 값들인

와

가 주어졌을 때 각 시간 단계에서 목표 값

를 예측하는 것이다.According to one embodiment, the dataset is

Consists of three patients,

The first patient

It can be assumed that it is expressed as a sequence of elements (or visits). In other words,

And the corresponding target value

to be. When modeling a patient, the goal of the predictive task is all previous input features and target values.

Wow

The target value at each time step given is

Is to predict.

개의 서로 다른 의료 코드를 가진다면,

는 각 의료 코드가

-방문 데이터에 나타나는지 여부를 가리키는 바이너리 상태를 암호화할 수 있다. 이 때, 목표는 이전의 모든 이력인

가 주어졌을 때 매 시간

마다

을 예측하는 것이다. 여기에서는 질병 진행 모델링(DPM)의 특수한 케이스라고 생각할 수 있는 진단을 위한 학습(L2D)에서, 방문 시퀀스의 가장 마지막에 특정한 질병을 진단한다.The formulation of this problem can incorporate several problems that arise in the analysis of electronic health records (EHRs), such as disease progression modeling (DPM) or Learning to Diagnose (L2D). In disease progression modeling (DPM), it is possible to simultaneously predict the evolution of the medical code at every point in time. Specifically, in the electronic health record (EHR)

If you have two different medical codes,

Each medical code has

-Binary state indicating whether it appears in the visited data can be encrypted. At this time, the goal is all previous history

Every hour when is given

each

Is to predict. Here, in learning for diagnosis (L2D), which can be considered a special case of disease progression modeling (DPM), a specific disease is diagnosed at the end of the visit sequence.

번째 환자에 대한 함수

를 다음과 같은 멀티-태스크 학습 패러다임에서 두 개의 함수인 공유 함수

와 개별 함수

로 분해하는 접근방식에 대하여 설명할 수 있으며, 다음 식과 같이 나타낼 수 있다.

For the first patient

In the following multi-task learning paradigm:

And individual functions

The decomposition approach can be explained, and can be expressed as the following equation.

[Equation 1]

는 임베딩 함수(embedding function)

를 통해 수행한 시간

에서 입력

에 대한 임베딩(embedding)이다.

는 전체적인 다양한 환자들의 전반적인 경향을 모델링하므로 이는 모든 환자들에게 공유될 수 있다. 반면,

는 공통적인(전반적인) 경향인 공통 함수

가 포착하지 않는 (

번째 환자에 대한) 환자-특정 신호를 모델링할 수 있다. 이 때, 특징 학습에 유용한 선택적인 임베딩을 제외하고는 개별 함수

를 통해서 환자들에 대해 어떠한 정보도 공유되지 않는다. 또한, 기존의 베이지안 통계(Bayesian statistics)에서 보통 파라미터 수준의 정보 공유 전략을 사용한다. 즉, 각 태스크에 대한 파라미터 벡터는 공유 파라미터 및 개별적인 파라미터의 합으로 표현될 수 있다. 그러나, [수학식 1]에서, 함수 값은 혼합된다. 두 접근 방식은 오직 공유 함수

와 개별 함수

가 선형 매핑일 때만 등가인데, 이는 일반적인 케이스가 아니다. 프레임워크 [수학식 1]의 그래픽 표현은 도 5와 같이 도시될 수 있다.here,

Is an embedding function

Time taken through

Input from

This is the embedding of

Models the overall trend of the various patients as a whole, so it can be shared among all patients. On the other hand,

Is a common function which is a common (overall) trend

Does not capture (

Patient-specific signals) for the first patient. In this case, individual functions except for optional embeddings useful for feature learning

No information is shared about patients through In addition, the existing Bayesian statistics usually use a parameter-level information sharing strategy. That is, the parameter vector for each task may be expressed as a sum of shared parameters and individual parameters. However, in [Equation 1], the function values are mixed. Both approaches are only shared functions

And individual functions

Is equivalent only when is a linear mapping, which is not a typical case. A graphical representation of the framework [Equation 1] may be illustrated as shown in FIG. 5.

5 shows a graphic representation of a multi-task effect framework according to an embodiment. Referring to FIG. 5, a graphic representation of the personalized healthcare multi-task effect framework of [Equation 1] can be confirmed.

A composite model apparatus using a recurrent neural network and a Gaussian process according to an embodiment is a mixed effect composite model of a recurrent neural network (RNN) and a Gaussian process.

와 개별 함수

는 고유의 원하는 성질을 가지고 있다. 특히, 개별 함수

를 위해 개인화된 가우시안 프로세스(personalized Gaussian Process)를 채택할 수 있으며, 이를 통해 비교적 적은 수의 데이터 지점(또는 특정 환자의 방문)을 기반으로 하여 (숨겨진 요인들을 통해) 확실하게 개별적인 신호들을 추정할 뿐만 아니라, 예측 불확실성을 자연적으로 제공할 수 있다. 또한, 개인화된 공식을 통해 많은 수의 환자를 다룰 때에도 많은 계산 비용을 피할 수 있는데, 이러한 점은 전체적인 모델을 확장 가능하게 만든다. 특히,

명의 환자가 있고 각 환자마다

번의 방문이 있을 때, 제안된 모델의 정확한 계산 비용은 많은 수의 환자를 다루지 못하는 완전한(full) 가우시안 프로세스(GP)의

가 아니라

이다.Shared function

And individual functions

Has its own desired properties. In particular, individual functions

A personalized Gaussian process can be adopted for this, which not only reliably estimates individual signals (via hidden factors) based on a relatively small number of data points (or visits of a specific patient). Rather, it can naturally provide predictive uncertainty. In addition, a large number of computational costs can be avoided even when dealing with a large number of patients through a personalized formula, which makes the overall model scalable. Especially,

There are 3 patients and each patient

With one visit, the exact computational cost of the proposed model is that of a full Gaussian process (GP) that does not handle a large number of patients.

Not

to be.

에서, MLP 또는 순환 신경망(RNN)과 같이 수식적으로 강력한 "심층 아키텍처(deep architecture)"를 사용할 수 있다. 여기서, MLP는 시그모이드(sigmoid) 활성화 기능을 가진 다중 계층 퍼셉트론 신경 회로망이다. 이는 고차원 의료 데이터의 복잡한 패턴을 포착하기 위한 합리적인 선택이며, 도 1에 도시된 바와 같이, 이를 사용하여 가우시안 프로세스(GP)의 복귀 문제(reverting problem)도 해결할 수 있다. 또한, 심층 모델(순환 신경망)을 이용하는 경우 전자 건강 기록(EHR)의 많은 수의 환자 데이터를 계산적으로 잘 다룰 수 있으며, 이는 모델을 확장 가능하게 만든다.Shared function

In, it is possible to use a mathematically powerful "deep architecture" such as MLP or recurrent neural network (RNN). Here, the MLP is a multi-layer perceptron neural network with a sigmoid activation function. This is a reasonable choice for capturing a complex pattern of high-dimensional medical data, and as shown in FIG. 1, the reverting problem of the Gaussian process (GP) can be solved by using this. In addition, the deep model (recurrent neural network) can handle large numbers of patient data in an electronic health record (EHR) computationally well, which makes the model scalable.

에 대해 다음과 같은 심층 글로벌 평균 함수(deep global mean function)를 공유하여 개인화된 가우시안 프로세스(GP)로 간소화될 수 있으며, 다음 식과 같이 나타낼 수 있다.The complex model [Equation 1] with these sophisticated options

It can be simplified to a personalized Gaussian process (GP) by sharing a deep global mean function as follows, and can be expressed as the following equation.

[Equation 2]

는

로 이름이 바뀌는데, 이는 가우시안 프로세스(GP)에 대한 평균 함수이다.

는 개별적인 프로세스인 개별 함수

로부터의 공분산 함수이다. 모델의 일부분이 결정론적 매핑(deterministic mapping)으로 표현되었지만 [수학식 2]의 그래픽 표현은 도 6과 같이 도시될 수 있다. 도 6은 일 실시예에 따른 순환 신경망과 가우시안 프로세스를 이용한 복합 모델(MECGP)의 그래픽 표현을 나타낸다.Here, the deep shared function

Is

Renamed to, which is the average function for the Gaussian process (GP).

Is an individual function that is an individual process

Is the covariance function from Although a part of the model is expressed by deterministic mapping, a graphic representation of [Equation 2] may be illustrated as shown in FIG. 6. 6 shows a graphical representation of a composite model (MECGP) using a recurrent neural network and a Gaussian process according to an embodiment.

7 shows a graphical representation of an individual Gaussian Process (GP) model for a patient in one embodiment.

Referring to FIG. 7, a graphical representation of an individual Gaussian process (GP) model for a patient in a composite model (MECGP) using a recurrent neural network and a Gaussian process according to an embodiment is shown.

가 구체적인 형태를 갖도록 제한하지 않는다. 그러나, 전자 건강 기록(EHR)의 순차적인 특징을 완전히 다루기 위하여 다음 식과 같이 순환 신경망(RNN)에 집중할 수 있다.The framework of [Equation 2] is an average function for a Gaussian process (GP)

Is not limited to have a specific shape. However, in order to fully address the sequential features of the electronic health record (EHR), one can focus on the recurrent neural network (RNN) as shown in the following equation.

[Equation 3]

는 각각 임베딩(embedding), 숨겨진 순환 유닛(hidden recurrent units) 및 출력(outputs)을 위한 (모든 환자들이 공유하고 있는) 순환 신경망(RNN)의 가중 파라미터이며, 이는 도 3과 같이 도시될 수 있다. here,

Is a weighting parameter of a recurrent neural network (RNN) (shared by all patients) for embedding, hidden recurrent units and outputs, respectively, which can be illustrated as in FIG. 3.

가 가우시안 프로세스라고 가정하여 제안된 프레임워크를 정규화(generalize)할 수 있다. 순환 신경망과 가우시안 프로세스를 이용한 복합 모델(MECGP)이라고 불리는 정규화된 프레임워크에서, 멀티-태스크 학습을 위한 기존의 많은 기초 모델이 이 프레임워크에서 실제적으로 표현될 수 있는지 검토할 수 있다. Furthermore,

Assuming that is a Gaussian process, the proposed framework can be normalized. In a normalized framework called composite model (MECGP) using recurrent neural networks and Gaussian processes, it is possible to examine whether many existing basic models for multi-task learning can be represented in this framework.

The following describes the optimization of a complex model (MECGP) using a recurrent neural network and a Gaussian process.

및

를 찾아낼 수 있다. 전자의 파라미터 집합

은 글로벌 공유 평균 함수(global shared mean function)에서 얻어지며, 후자의 파라미터 집합

은 개인화된 가우시안 프로세스(GP)에서 얻어진다. 제안된 모델은 회귀(regression) 및 분류(classification) 태스크 모두에 범용적으로 적용 가능하기 때문에 최적화 절차를 추상적으로 제시한다. 이 때, 학습 목표는 다음 식을 최소화하는 것이다.The learning goal is to minimize the sum of the set of negative marginal log-likelihoods of the Gaussian process (GP) model obtained from the learners, which is a set of two parameters.

And

Can be found. Former parameter set

Is obtained from the global shared mean function, the latter parameter set

Is obtained in a personalized Gaussian process (GP). Since the proposed model is universally applicable to both regression and classification tasks, the optimization procedure is presented abstractly. At this time, the learning goal is to minimize the following equation.

[Equation 4]

는 글로벌 평균 함수의 일반화(regularization)를 위한 하이퍼 파라미터이고,

는 최소화를 위한 개별적인 가우시안 프로세스(GP)의 손실이다. 가우시안 가능도(Gaussian likelihood)의 경우,

는 개별적인 가우시안 프로세스(GP)의 음경계 로그-가능도이며, 다음 식과 같이 나타낼 수 있다. here,

Is a hyperparameter for regularization of the global average function,

Is the loss of the individual Gaussian process (GP) for minimization. For Gaussian likelihood,

Is the penile system log-likelihood of an individual Gaussian process (GP), and can be expressed as the following equation.

[Equation 5]

는 각 환자에 대한

가 주어졌을 때의 완전 공분산 행렬(full covariance matrix)이다. 비가우시안 가능도(non-Gaussian likelihood)의 경우, 손실은 변분 근사(variational approximation)에 의한 변분 하한 경계(variational lower bound)로 계산되거나 시뮬레이션에 의할 수 있다(비특허문헌 2). 두 개의 파라미터 집합에 대한 [수학식 4]의 그래디언트(gradient)는 다음과 같이 체인 규칙(chain rule)에 의해 유도될 수 있으며, 다음 식과 같이 나타낼 수 있다.here,

Is for each patient

Is the full covariance matrix given In the case of non-Gaussian likelihood, the loss can be calculated as a variational lower bound by a variational approximation or can be by simulation (Non-Patent Document 2). The gradient of [Equation 4] for the two parameter sets can be derived by a chain rule as follows, and can be expressed as the following equation.

[Equation 6]

의 그래디언트(gradient) 계산은 추가적인

항을 포함하며, 이는 조금 더 복잡한 계산을 야기한다.Here, parameter dependency and regularization terms are not used for the sake of simplicity of expression. Unlike in the case of a vanilla recurrent neural network (RNN), the parameter set in [Equation 6]

The gradient calculation of

Term, which leads to slightly more complex calculations.

The overall algorithm for learning parameters in an alternative way can be summarized as follows.

[Table 1]

Referring to Table 1, Algorithm 1 shows the optimization of a complex model (MECGP) using a recurrent neural network and a Gaussian process.

By setting up a personalized formulation, the high computational cost of the Gaussian process (GP) can be reduced. In addition, the deep architecture as a shared mean function can be efficiently updated through a general back-propagation algorithm.

In Equation 2, the inference of the individual Gaussian process (GP) follows the general procedure of the Gaussian process (GP) inference. For simplicity, we only describe how the final prediction is made through Gaussian likelihood. The exact same approximation techniques developed for the classification problem using the Gaussian process (GP) can be applied perfectly to this case as well.

에 대해

를 예측하려 한다고 가정한다. 명료성을 위해 데이터에서 환자 인덱스

가 제거된 대응하는 데이터

및

가 주어진다. 우선 개인화된 추론을 위해 새로운 가우시안 프로세스(GP)를 둔다. 이후, 시계열 입력 특징을 임베딩하기(embed) 위하여 모든

에 대해

라고 표현되는 순환 신경망(RNN)의 모든 숨겨진 상태를 계산한다. 숨겨진 상태 및 대응하는 목표 값을 사용해서, MLE에 의한 새로운 가우시안 프로세스(GP)의 파라미터를 갱신하는데, 최적화 프로세스에서 수행했지만 평균 함수가 고정되었기 때문에 이를 적응 프로세스(adaptation process)라고 부른다.

에 대한 예측 분포(predictive distribution)는

이며, 다음 식과 같이 나타낼 수 있다.New patient

About

Suppose you want to predict Patient index in data for clarity

Corresponding data has been removed

And

Is given. First, we put a new Gaussian process (GP) for personalized reasoning. After that, to embed all the time series input features

About

Compute all hidden states of the recurrent neural network (RNN), expressed as. Using the hidden state and the corresponding target value, the parameters of the new Gaussian process (GP) are updated by the MLE, which is called the adaptation process because it was performed in the optimization process but the average function is fixed.

The predictive distribution for is

And can be expressed as the following equation.

[Equation 7]

이며

이다. 예측은 순차적인 방법으로 수행될 수 있으며, 이는 적응 프로세스를 통하여 어느 환자의 임의의 시간 지점에 대한 출력을 예측할 수 있다는 것을 의미한다. 첫 번째 시간 지점

에서의 예측은 글로벌 평균 함수(global mean function)에 의해 결정론적으로(deterministically) 수행될 수 있고 이는 평균적인 예측이라는 것을 의미한다. 시간 지점

를 증가시킬수록 환자에 대한 더 많은 증거를 가질 수 있고 개선된 개인화된 예측을 할 수 있다.here,

Is

to be. The prediction can be performed in a sequential manner, which means that the output for any time point of any patient can be predicted through the adaptation process. First time point

The prediction at can be performed deterministically by a global mean function, which means that it is an average prediction. Time point

The more you increase the, the more evidence you have about the patient and you can make improved personalized predictions.

는

에 따라 적절히 정의된 어떤 분포

를 따른다(즉,

일 때는 정규 분포이며,

일 때는 베르누이 분포(Bernoulli distribution)이다). 그 후, 시험 케이스에 대한 가우시안 프로세스(GP)의 잠재 함수(latent function)의 분포

는 다음 식과 같이 주어질 수 있다.As the loss described above, approximate predictions can be derived using the non-Gaussian likelihood of the classification problem. The expression used in the Gaussian likelihood described above, and the output

Is

Any distribution properly defined according to

Follows (i.e.

Is a normal distribution,

When is, it is a Bernoulli distribution). Then, the distribution of the latent function of the Gaussian process (GP) for the test case

Can be given as the following equation.

[Equation 8]

이다. 마지막으로,

에 대한 예측 분포(predictive distribution)는 다음 식과 같이 나타낼 수 있다.Here, by Baye's rule

to be. Finally,

The predictive distribution for can be expressed as the following equation.

[Equation 9]

는 문제의 종류에 따라

가 주어졌을 때

에 대해 적절히 설계된 가능도 함수이다. 회귀(regression)의 경우, 위에서 제시한 바와 같이,

가 가우시안을 따를 때 [수학식 8] 및 [수학식 9]에 대해 분석적 형태를 취한다. 반면, 분류(classification) 문제의 경우, 가능도 함수는

과 같은 시그모이드(sigmoid) 함수로 설계되며, 이는 [수학식 8] 및 [수학식 9]에서의 적분을 분석적으로 아주 다루기 힘들게 만든다. here,

Depending on the type of problem

When is given

Is a properly designed likelihood function for. In the case of regression, as suggested above,

When g follows Gaussian, it takes an analytic form for [Equation 8] and [Equation 9]. On the other hand, in the case of the classification problem, the likelihood function is

It is designed with a sigmoid function such as, which makes the integral in [Equation 8] and [Equation 9] analytically difficult to handle.

에 대한 근사 방법을 필요로 한다(비특허문헌 2).Thus, post-mortem approximation such as Laplace approximation, variational method, or Markov chain Monte Carlo (MCMC) approximation

It requires an approximation method for (Non-Patent Document 2).

As described above, the embodiments are machine learning models capable of providing reliable and personalized prediction in the medical field as a complex model combining a recurrent neural network and a Gaussian process. Data-based machine learning models are showing excellent results in many fields based on vast amounts of data, and a field that has recently been in the spotlight among machine learning fields is a deep learning (deep learning) technique using artificial neural networks. In the medical field, many attempts have been made using deep learning techniques to provide high-quality and efficient medical services based on electronic health record (HER) data in which all patient information is recorded. However, it is not easy to apply deep learning techniques in the medical field compared to other fields, and the biggest reason is that these techniques cannot provide uncertainty about prediction.

Accordingly, the embodiments complemented the shortcomings of deep learning by combining a Gaussian process model capable of providing prediction uncertainty to a recurrent neural network. In addition, by proposing a framework for creating personalized models for each patient, it is possible to efficiently model patient characteristics hidden in the data and provide personalized predictions for each patient characteristic. This model proved its excellence through comparison with general recurrent neural networks and other modern machine learning models using real medical data.

The apparatus described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices and components described in the embodiments include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It can be implemented using one or more general purpose computers or special purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to behave as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. Can be embodyed. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of the program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described by the limited embodiments and drawings, various modifications and variations are possible from the above description by those of ordinary skill in the art. For example, the described techniques are performed in a different order from the described method, and/or components such as a system, structure, device, circuit, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and claims and equivalents fall within the scope of the claims to be described later.

Claims (12)

A recurrent neural network (RNN) that acquires a shared characteristic, which is an overall trend among multiple heterogeneous patients from time series data; And
Gaussian Processes (GP) modeling individual characteristics for each patient from the time series data
Including,
For multi-task learning of the recurrent neural network and the Gaussian process, a function consisting of time series data for each patient is decomposed into a sum of a shared function and an individual function, and the shared function is used to obtain a shared characteristic using the recurrent neural network. To train the recurrent neural network, and the individual function trains the Gaussian process to model individual characteristics using the Gaussian process,
Acquiring the shared characteristic and the individual characteristic when inputting new time series data through the learned recurrent neural network and the Gaussian process
A complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 1,
Optimizer for minimizing the sum of the set of negative marginal log-likelihood of the Gaussian process obtained from the patients
A complex model device using a recurrent neural network and a Gaussian process further comprising. delete The method of claim 2,
The optimization unit,
Minimizing the sum of the set of negative marginal log-likelihoods using the shared characteristics obtained from the recurrent neural network and the individual characteristics obtained from the Gaussian process
A complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 1,
The time series data,
What is Electronic Health Care (EHR)
A complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 1,
A single Gaussian process model is constructed for each patient by acquiring shared characteristics through the recurrent neural network and reflecting individual characteristics through the Gaussian process, and the single Gaussian process model is a personalized prediction for each patient's characteristics. To provide
A complex model device using a recurrent neural network and a Gaussian process, characterized by. Obtaining a shared characteristic, which is an overall tendency among heterogeneous patients, from time series data using a recurrent neural network (RNN); And
Modeling individual characteristics for each patient from the time series data using Gaussian Processes (GP)
Including,
For multi-task learning of the recurrent neural network and the Gaussian process, a function consisting of time series data for each patient is decomposed into a sum of a shared function and an individual function, and the shared function is used to obtain a shared characteristic using the recurrent neural network. To train the recurrent neural network, and the individual function trains the Gaussian process to model individual characteristics using the Gaussian process,
Acquiring the shared characteristic and the individual characteristic when inputting new time series data through the learned recurrent neural network and the Gaussian process
A method of operating a complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 7,
Optimization step of minimizing the sum of the set of negative marginal log-likelihoods of the Gaussian process obtained from the patients
A method of operating a complex model apparatus using a recurrent neural network and a Gaussian process further comprising a. delete The method of claim 8,
The optimization step,
Minimizing the sum of the set of negative marginal log-likelihoods using the shared characteristics obtained from the recurrent neural network and the individual characteristics obtained from the Gaussian process
A method of operating a complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 7,
The time series data,
What is Electronic Health Care (EHR)
A method of operating a complex model device using a recurrent neural network and a Gaussian process, characterized by. The method of claim 7,
A single Gaussian process model is constructed for each patient by acquiring shared characteristics through the recurrent neural network and reflecting individual characteristics through the Gaussian process, and the single Gaussian process model is a personalized prediction for each patient's characteristics. To provide
A method of operating a complex model device using a recurrent neural network and a Gaussian process, characterized by.

Images