KR20070020463A - System and method for the automatic generation of hierarchical tree networks using two supplementary learning algorithms optimized for each leaf of the hierarchical tree network - Google Patents

Abstract

A system and method for creating a hierarchical tree network and using a linear-plus-nonlinear algorithm to form a consensus on the future health status of a member. Each leaf in the hierarchical tree network is uniform in clinical characteristics, experience cycle, and valid data assets. Optimization is performed on each leaf so that features and learning algorithms can be adapted to the local characteristics specified for each leaf.

Description

SYSTEM AND METHOD FOR AUTOMATIC GENERATION OF A HIERARCHICAL TREE NETWORK AND THE USE OF TWO COMPLEMENTARY Using two complementary learning algorithms optimized for each leaf of the hierarchical tree network LEARNING ALGORITHMS, OPTIMIZED FOR EACH LEAF OF THE HIERARCHICAL TREE NETWORK}

This application describes the use of two supplementary learning algorithms optimized for each leaf of a hierarchical tree, cited herein, and a system and method for automatic generation of a hierarchical tree network. The filed US claims the benefit of patent application serial number 60 / 565,579.

The present invention relates to a system and method for creating hierarchical tree networks, using a linear-plus-nonlinear learning algorithm to form a consensus view of a member's future health status. Each leaf in the hierarchical tree network is uniform in clinical characteristics, experience cycle, and valid data assets. Optimization is performed on each leaf so that features and learning algorithms can be adapted to the local characteristics specified for each leaf.

The present invention is for a method of predicting a future health condition of an individual. The method is:

a. Establishing a hierarchical tree network assigning each member only to at least one of the plurality of nodes as a function of a plurality of stratification variables for the plurality of members;

b. Providing, to a computer-based system, member demographic data, active member medical billing data, and active member medical billing data for a plurality of members and each member;

c. For each node, to identify an optimal subset of features from a set comprising at least one of the member demographic data, active member medical claim data, and active member medical claim data for all members assigned to the node. Performing feature selection on each of the plurality of nodes;

d. Training an MVLR algorithm and a BRN algorithm utilizing one or more of the member demographic data, active member medical billing data, and active member medical billing data, and storing the learned parameters in a database to create a learned parameter database. Doing;

e. Using the learned parameter database, using member demographic data, active member medical billing data, and active member medical billing data of one or more members for one or more members, and using the MVLR algorithm to obtain MVLR future health status scores. Calculating and calculating a BRN future health status score using a BRN algorithm, and determining a final score by calculating an arithmetic mean of the MVLR future health status score and the BRN future health status score.

Stratification variables may be based on the member's registration period, for example 6 months or less; Presence of catastrophic conditions; Presence of diabetes; Presence of a drug trigger or an inpatient admission trigger; And the presence of medical and medical billing data, the presence of medical billing data only or the presence of medical billing data only.

The method may also include a reporting step in which information about the member is presented with the calculated final score. For example, member information includes enrollment / eligibility information, clinical status information including status, trigger type, and trigger date, and representation of member costs over time for medication and / or medical claims. can do.

The invention will be more readily understood by reference to the following description taken in conjunction with the accompanying drawings.

1 illustrates an N-dimensional vector space that has been processed and mapped to support the description of feature optimization and learning, along with the relative advantages and vulnerabilities of various learning algorithms.

2 shows an overall flow diagram of hierarchical tree network generation, feature optimization, learning and scoring.

3A-3H illustrate two selected learning algorithms (in combination of multivariate linear regression and Bayesian regularization network) in action.

4A and 4B show glymps of memory for multilayer perceptrons.

5 is a different answer represented by a myriad scatter points away from the line indicating perfect agreement while the two learning algorithms, base normalization network (BRN) and multivariate linear regression (MVLR) generally agree in some cases. to provide an answer.

6A-6F show plots of the future cost in one year versus the square root of various covariates. These simple plots provide many useful insights. If a simple linear relationship exists, a monotonic increasing line of straight lines will appear. Unfortunately, none of these features show this clear trend, indicating that there is no simple linear insight.

As a background, predictive models were used in the health-insurance industry to predict the future health status of members. The model can be adapted for the actuarial underwriting and identification of high-risk members for proactive clinical intervention.

Constructing a predictive model involves two stages, first converting billing data into a set of clinical features, and second, using historical billing data to determine the relationship between clinical features (x) and future health conditions (y). To learn. Typical clinical features include binary clinical markers. Note that IHCIS uses hundreds of binary clinical markers as covariates in predictive models. Examples of binary markers include inpatient episodes in the last three months, coronary artery bypass graft (CABG) procedures for the previous year, heart attack, age, gender, Rx costs for previous years, medical costs for the previous year, length of hospital stay, number of ER visitors, and Rx-to-med rates.

Currently adopted learning algorithms include rule-based expert systems (Active Health), linear regression (IHCIS, Ingenix, DxCG, Active Health and Biosignia), multi- Layer perceptrons (Mckesson, classification-and-regression trees (CART) (Medical Scientists), and k -nearest or discriminant adaptive near neighbor (Hastie) and Tibshirani, 1996) (MedAI), which cover variables of instance-based learners: the learning algorithms are L-1 (absolute error), L-2 (mean-squared error), and L-∞ (maximum). Error) We try to find the relationship between input and output by optimizing an objective function governed by various types of error terms, such as norm.

Conceptually, feature extraction and optimization facilitates learning by finding the smallest feature dimension that can adequately capture all of the useful information present in the raw data. Learning is similar to finding the mapping function between the optimal feature and the desired output. Learning algorithms can be broadly categorized into parametric, nonparametric and boundary-decision types (Kil and Shin, 1996).

Parametric algorithms make strong parametric assumptions about data distribution such as linear Gaussian. A simple parametric hypothesis with a correspondingly small number of model parameters for adjustment may be a double-edged sword. If the actual data distribution is much more complex than the original assumption, the parametric learner will experience a large model mismatch, which occurs when the learning algorithm is inadequate for capturing high nonlinear relationships between x and y. Paradoxically, in the presence of large data mismatches that often occur when real-world data is somewhat different from the modeled training data, a simple parametric algorithm looks at its nonparametric counterparts. Tends to be excellent.

Non-parametric and boundary-decision learning algorithms attempt to learn data distribution from the data. In general, neural networks and non-parametric learning algorithms are good at minimizing model mismatch errors, but tend to be difficult and ambiguous. There is no simple way to describe the relationship between input and output. Since this simple linear relationship does not exist, it cannot be said that higher previous year's medical costs result in higher future costs. These learning algorithms have a large number of adjustable parameters that can help capture complex and nonlinear relationships. Unfortunately, such fine-tuning can cause overfitting or memory that can be detrimental in the presence of data mismatches. It is also known that nonlinear algorithms tend to produce wild results when operating outside the trained space. That is, this is ill-suited when a large discrepancy between the modeled training data and the real world data is expected.

The Hidden Markov model is useful for modeling temporal transitions. In practical applications it is often observed that computationally expensive Baum-Welch optimization is inferior to segmental k-mean (SKM) despite the higher reported accuracy. The reason is the trade off between model and data mismatch. In most real world problems we have seen so far, data mismatch is much more important than model mismatch.

From left to right, to demonstrate the core concept of feature optimization and learning, FIG. 1 shows an M-dimensional vector space in which multiple overlaps are signaled and spanned by good features determined after feature optimization. Shows the original N-dimensional vector space spanned by the raw data whose features are ranked to produce the equation, where M << N, and the mapping on the right is the final 1- generated by the classifier. Show the dimensional decision space. The following summarizes the relative strengths and weaknesses of the various learning algorithms. The characteristic of a parametric learning algorithm is that it makes a strong parametric assumption about the underlying classification-state probability distribution, very simple to train, and prone to model mismatch. Examples of parametric learning algorithms are multivariate Gaussian classifiers, Gaussian mixed models, and linear regression. The nature of non-parametric learning algorithms is that they do not make parametric assumptions, learn distribution from data, are expensive to train in most instances, and are prone to data mismatch between training and test data sets. Examples of non-parametric learning algorithms are Kernel estimator, histogram, functional form, and K-near list. The property of a boundary decision learning algorithm is that it constitutes a (non) linear boundary function that separates multiple classes, and is very expensive to train, and that internal parameters are determined heuristically in most instances. Examples of boundary decision learning algorithms include multi-layer perceptrons, discriminant neural networks, and support vector machines.

을 이용하여 y를 추정하며, 여기서, a_n은 x_n와 y 간의 상관관계의 강도 및 방향을 나타낸다. 선형 회귀의 직관적 성질을 나타내게 된다. 전년도 의료 비용에 대한 회귀 계수가 +1.3인 경우, 다른 모든 입력 변수들이 동일 하다고 한다면 향후 비용은 주어진 전년도 의료 비용의 1.3 배가 될 것이라고 추측할 수 있다. The popularity of linear regression arises from its intuitive power. Let y be the future cost, and x be the clinical feature or covariate ( ∈ R ^N ). Linear regression has the following equation:

And y is estimated, where a _n represents the strength and direction of the correlation between x _n and y. Indicate the intuitive nature of linear regression. If the regression coefficient for medical costs for the previous year is +1.3, then all other input variables are equal, it can be assumed that future costs will be 1.3 times the given previous year's medical costs.

Despite its ease of interpretation, linear regression cannot model complex nonlinear relationships due to its simple mathematical formulation that can cause large model-mismatch errors. Nevertheless, it is an unanswered question how to combine the strengths of linear and nonlinear models to minimize both data- and model-mismatch errors.

In summary, the success of modeling depends on the trade-off between model mismatch and data mismatch. It is imperative to find an integrated and complementary set of algorithms, extraction of derived features, optimization and robust learning in data transformation.

Instead of trying to solve the whole problem in one broad stroke, the present invention relies on a judicious divisional conquest approach. First, the problem space is divided into logical subspaces using a hierarchical tree network. Each leaf of the hierarchical tree network is uniform in clinical characteristics, experience cycle, and valid data assets. Moreover, hierarchical tree construction methods can include clinical status scores that represent the total disease burden of members, prior-experience characteristics (monthly costs), chronic versus overall costs, and disease tragectory (increasing, Flexible enough to accommodate additional dimensions such as reduction or constant over time). The next feature ranking is done at each leaf to take advantage of locally specific features useful for prediction. Finally, the majority learning algorithm uses its own thinking hat to examine the relationship between the output and the optimal subset of features. 2 illustrates one embodiment of a divide-and-conquer approach.

Hereinafter, each step will be described in detail.

, 여기서, b _k (m)은 m 번째 구성원에 대한 k 번째 존재/부재 플래그이고, pppm(k)은 k 번째 동반질환 질병 세트를 갖는 모든 구성원에 대한 평균 매월 비용을 나타낸다. 이 2-차원 벡터 공간에서, 1 사분면은 PMPM(per-member-per-month) 이전 비용을 갖는 치명적으로 아픈 개체군(population)을 나타낸다. 3 사분면은 낮은 PMPM을 갖는 비교적 건강한 개체군을 나타낸다. 2 사분면은 비교적 적은 수의 치명적 임상 상태를 가지나 큰 PMPM(아마도 양태적 이슈(behavioral issue))을 갖는 구성원들을 포함한다. 최종적으로, 4 사분면은 무수한 치명적 임계 상태를 가지나, 아마도 구성원들이 그 상태를 더욱 양호하게 자기-관리하기 때문에 작은 PMPM을 갖는 구성원들을 나타낸다. Hierarchical Tree Network Creation: Design a hierarchical tree network, where each leaf represents a clinically uniform cluster that shares similar huma or humana age and data flags. Clustering will be performed in 2-D space for the Clinical Condition Score (CCS) on the x-axis and the monthly transfer cost on the y-axis. The clinical status score for the mth member with multiple comorbid conditions is defined as follows:

Where b _k (m) is the k th presence / absence flag for the m th member, and pppm (k) represents the average monthly cost for all members with the k th comorbidity disease set. In this two-dimensional vector space, quadrant 1 represents a fatally ill population with the cost of per-member-per-month transfer. The third quadrant represents a relatively healthy population with low PMPM. The second quadrant includes members with a relatively small number of fatal clinical conditions but with a large PMPM (possibly a behavioral issue). Finally, the fourth quadrant has a myriad of deadly critical states, but represents members with small PMPMs, perhaps because they better self-manage that state.

a. Guide to tree subdivision: pre- and post-entropy reduction measures (ie, what do you get by partitioning?), Level of data mismatch (full-data-set) set) MVLR performance difference between training and ten-fold cross validation), minimum population size and need of 300 (registration period and data validity).

b. Branch recombination: Recombination with nodes in different branches based on similarity metrics such as Kullback-Leibler divergence is defined as follows:

Where p (y | leaf = j) is the probability of the selected output associated with the hierarchical tree leaf j.

c. Approximation: Given the reality of our data, we use the following five features to create hierarchical tree leaves.

1. Humana membership duration: This indicates how long a member has been with Humana and determines the amount of valid claim history. This field is necessary to take into account the high turnover rate business feasibility in the health insurance industry.

2. Catastrophic status: This field represents the presence of expensive chronic conditions that require coordinated care management and rigorous clinical intervention.

3. Diabetes Flags: The Personal Nurse (PN) program focuses on behavioral modifications for diabetics with other chronic conditions, members who have diabetes but whose catastrophic status is not flagged separately.

4. Trigger Type: This field shows why members are led to a predictive modeling queue. Members with inpatient seizures need special exchange counseling and a reminder message about how they can be excluded from further hospitalization. In addition, they accept more for modal intervention messages.

5. Data validity: It goes without saying that models should consider valid data assets.

Feature subset selection: For each leaf, we perform feature optimization to select the optimal subset of features at the diminishing return point.

이다.a. Feature Correlation: To minimize redundancy and feature dimensions, we combine the highly correlated features (ρ≥0.9) using principal component analysis, where

to be.

b. Feature ranking: The purpose of feature ranking is to investigate the contribution of each feature to the overall prediction accuracy. If features are completely orthogonal (ie ρ _ij = 0, i ≠ j), feature ranking is based on Fisher's discriminant ratio, multi-modal overlap (MOM), divergence, batacha Many suitable metrics, such as Bhattacharyya distance (Kil and Shin, 1996), are used to degrade to marginal or one-dimensional feature rankings. If the features are not orthogonal, we can use a stochastic or combinatorial optimization algorithm.

Learning: Learning can take the form of regression (continuous dependent variables such as future costs) or classification (discrete dependent variables such as identifying the top 20% of future costs). The dependent variable can be oriented clinically or actuarially. We use the following learning algorithm:

.a. Multivariate Linear Regression (MVLR):

.

b. Base Normalization Network (BRN): BRN can be used for both regression and classification (Foresee and Hagan, 1997).

Scoring : Using a trained parameter database, score unknown population batches using multiple learning algorithms.

Now look at the learning and scoring of this embodiment as shown in FIG. 2. In the case of Figure 2, the following reference numbers are used: 1-As input of fused features, they are all valid medical and medical data and demographic data; 2-feature subset selection; 3-learning or scoring ?; 4-learning path; 5-MVLR learning; 6-BRN learning; 7-trained database; 8-score path; 9-MVLR scoring; 10-BRN scoring; 11-mean operator; 12-summary of clinical status; And 13-final score and status report.

Prior to the detailed description, the following table provides details about 30 nodes in a hierarchical tree network.

In the table above, the leaf identified by 1 to 15 forms a first set of nodes of members whose registration is less than 6 months, and the leaves 16 to 30 form a second set of nodes whose registration ≧ 6 months. . For the first set of nodes 1 to 15, the leaves 1 to 3 form a first subset and the leaves 4 to 15 form a second subset; Leaves 4 to 6 form a first subset, leaves 7 to 9 form a second subset, leafs 10 to 12 form a third subset, and leaves 13 to 15 ) Forms a fourth subset. For a second set of nodes 16-30, leaves 16-18 form a first subset, and leaves 19-30 form a second subset; Leaves 19 to 21 form a first subset, leaves 22 to 24 form a second subset, leaves 25 to 27 form a third subset, and leaves 28 to 30 ) Forms a fourth subset.

1. As shown above, in order to handle all situations where a predictive model can be considered, 30 nodes or leaves are derived as we create a hierarchical tree network as a function of the following five stratification variables. :

Enrollment Period: The length of time a member has with Humana. As shown above, the members are divided into members with Humana for 0- to 0.5 years and members with Humana for more than 0.5 years.

Catastrophic status: Does the member have a strict clinical condition? If the member has any of the following, the category flag is yes, otherwise no.

cancer.

Terminal illness (renal failure or liver failure).

transplantation.

Rare disease.

HIV.

CAD + CHF + hypertension.

Diabetes flag, yes or no. If the members have a catastrophic state in b above, they are on the leaves 1 to 3 or 16 to 18 and the diabetic flag has no effect.

Trigger Type: Allow new Rx billing or inpatients.

Valid data: Rx only, medical only, or both.

As an example, from the above table and the above description of 30 nodes, we say that if John Smith had been with Humana for 9 months and had a catastrophic state, he would have medicinal and medical benefits, he would reef You will be at # 16. On the other hand, if Nancy Doe has been with Humana for 12 months and does not have any catastrophic status but is diabetic and has an impact trigger, assuming she has a health and medical benefit plan guaranteed by Humana, The draw will be on leaf # 22. In addition, only members having a catastrophic state or having an admission admission trigger or a subscription trigger will be assigned to one of the 30 nodes.

For each leaf, we perform feature optimization to find the optimal subset using add-on combinatorial optimization.

In training, we train both multivariate linear regression (MVLR) and base normalization network (BRN) algorithms. The learned parameters are stored in the local database. Two learning algorithms are selected to cope with data-mismatch and model-mismatch errors typically encountered in real-world conditions. Model-mismatch errors occur when the learning algorithm is too simple to properly model the complex relationship between input and output. Data-mismatch errors, on the other hand, occur when a supertuned learning algorithm cannot cope with differences in data characteristics between training data and test data.

In scoring, two learning algorithms provide their assessment of a member's future health status.

A simple arithmetic mean operator outputs the final score, which is then combined with the summary of clinical status of the members in the report format. The clinical status report includes up to five MCC status, trigger type, trigger date, scoring date, personal key, two codes for insulin, oral or diabetes, and a maternity flag. It is composed.

For each leaf described above, separate learning is done for the structurally uniform population segments belonging to that leaf. For example, we get the following statistics for the leaves 16-18:

에서 측정된 성능, 여기서 y 및

는 각각 실제 및 예측 된 출력을 나타내며, 바(var)(ㆍ)는 분산 연산자를 나타낸다:One.

Performance measured at, where y and

Represent the actual and predicted output, respectively, and var (·) represent the variance operator:

  a. Rx + med: 0.41 / 0.39 (BRN / MVLR)

  b. Rx only: 0.24 / 0.18

  c. Med only: 0.36 / 0.35

2. The top five features are:

  a. Rx + med

i. Square root of the total prior paid amount

ii. For other inpatient facilities

Mean utilization

iii. Trend of previous payments over time

iv. Variation in Payments-The more consistent, the easier the forecast.

v. Multiple sclerosis, hemophiliac

Costs associated with rare diseases such as

  b. Rx only

i. Nonzero Rx Cost Term

ii. Number of unique drugs adopted

iii. Ace drug cost

iv. Ratio of Rx chronic cost to Rx total cost

-The higher the ratio, the easier the prediction

v. GPI drug class id

  c. Med only

i. Square root of payment last 6 months ago

ii. Average use of other outpatient facilities

iii. Number of Distinctive Diagnosis in Primary ICD-9

iv. Major Clinical Condition (MCC) Cost Averages per Month

v. Average payments to inpatient physicians

  3. In scoring, we compute a large number of features for each member, find out the leaf # to which she belongs, filter the subset of features we need for the selected leaf, and learn two about that leaf. Load the learning parameters associated with the algorithm and generate two outputs from MVLR and BRN. For consensus between the two learning algorithms we take an arithmetic mean. For MVLR, the learning parameters consist of normalization and regression coefficients. For BRN, the learning parameters include network architecture structure, optimization function, and weight / bias values for each network node and connection.

4. As an output, we append each individual clinical state, weather map, trigger information, and other eligibility information to the mean score. The member weather map provides a representation of member cost over a specified time period for the member's medication and / or medical claim. Mapping ICD9 codes for major clinical conditions can lead to different major clinical status categories (eg, coronary artery disease, congestive heart failure, other heart failure, circulatory system, cancer), (E.g., major clinical status coronary artery disease is subclasses of coronary artery bypass graft, percutaneous transluminal coronary angioplasty, myocardial infarction, angina, Medical cost information can be shown for that subset and subset of other ischemic heart diseases, coronary atherosclerosis, and hyperlipidemia. In addition, drug codes are mapped to various major clinical conditions and their subsets and subsets. In addition, the member weather map can be used for billing member usage and medical billing by treatment sites over the same specified time period (eg, hospital inpatient, hospital emergency room, physician office visit). Can be represented.

Next, we provide additional algorithmic details.

Bass Normalization Network:

Traditional feedforward neural networks repeatedly learn highly nonlinear relationships between inputs and outputs using multiple feedforward-connected neurons that perform weighted arithmetic summation and nonlinear activation. Unfortunately, performance has been found to lack robustness in the presence of data mismatches, often known as "new" inputs. This is the case when normalization plays an important role.

Normalization means finding a meaningful approximate solution, not finding the exact meaningless (Neumaier, 1998). A popular method of normalization in linear algebra is the so-called diagonal loading, also known as Tikhonov normalization.

이다. 의사 역행렬(pseudo inverse)(A ⁺ = ( A'A ) ^-1 A')이 존재하지 않는 풀 수 없는(ill-posed) 문제의 경우(이는 SVD(singular value decomposition)를 이용하여 이를 입증할 수 있음), 의미 있는 양호한 근사 솔루션은

이며, 여기서, 우리는 정규화를 위해 각각의 대각 항(diagonal term)에 소량을 추가한다. 이 피처는 우리의 다변 가우시안 분류 알고리즘이 된다. If we want to find solution x at y = Ax, the well-known standard L-2 solution

to be. In the case of ill-posed problems where a pseudo inverse ( A ⁺ = ( A'A ) ^-1 A ' ) does not exist (which can be demonstrated using SVD (singular value decomposition)). A good approximation solution

Where we add a small amount to each diagonal term for normalization. This feature becomes our multivariate Gaussian classification algorithm.

의 형태를 취하며, 여기서, D 및 S는 각각 데이터베이스 및 모델 구조를 나타낸다. N은 데이터의 샘플 크기인 한편, d는 주어진 모델 구조와 연계된 파라미터들의 세트를 나타내는

를 갖는 모델 S의 파라미터의 개수이다. 직관적으로, 이 방정식은 모델 복잡성을 최소화하는 동시에 데이터로부터 학습된 모델 구조의 설명력(explanation power)을 최대화함으로써 목적 함수가 최대화될 수 있다는 것을 나타낸다. Similar to learning, instead of using a traditional objective function that minimizes the mean squared error (L-2 norm), the BEI (Bayesian Information Criterion) or minimum description length (Rissanen, 1989) You can use conceptually similar optimization functions, which

Where D and S represent the database and model structure, respectively. N is the sample size of the data, while d represents the set of parameters associated with the given model structure

The number of parameters of model S with Intuitively, this equation indicates that the objective function can be maximized by maximizing the explanation power of the model structure learned from the data while minimizing model complexity.

이며, α+β=1이다. J의 첫번째 항은 잘 알려진 평균 제곱 예측 오차인 한편, 두번째 항은 네트워크 가중(network weight)의 제곱의 합을 나타낸다. β>>α이면, 학습은 예측 오차를 최소화하여 결과적인 네트워크 응답이 다른 것보다 훨씬 더 평활(smooth)하게 함으로써 가중 감소(weight reduction)를 훨씬 더 강조할 것이다. 가우시안 가정 하에서, α 및 β에 대해 더 근사한 형태의 솔루션을 도출하기 위해 베이스 룰(Bayes' rule)을 적용할 수 있다(Foresee and Hagan, 1997).In BRN, the objective function is

And α + β = 1. The first term of J is the well-known mean square prediction error, while the second term represents the sum of the squares of the network weights. If β >> α, learning will further emphasize weight reduction by minimizing the prediction error so that the resulting network response is much smoother than others. Under the Gaussian assumption, Bayes' rule can be applied to derive a more approximate form of solution for α and β (Foresee and Hagan, 1997).

Briefly, BRN relies on the use of base normalization to avoid overfitting during learning while maintaining the flexibility to learn and respond to new patterns. In this case, we use three factor-errors, network weighting, and model complexity in terms of the actual function dimension for greater robustness.

Linear regression:

는 다음과 같이 구현될 수 있다:function

Can be implemented as follows:

Again, we use Tihonov normalization to find a robust solution to the inverse matrix problem.

Next, we summarize how to embark on a robustness issue.

1. Objective function to address prediction error, model complexity, and model parameters. 2. Deformation of the dependent variable to reduce dynamic range and to scale out the square root of the cost instead of the linear outlier, i.e., the linear cost. 3. Performance analysis, including 5- or 10-fold cross validation and full-data-set training. 4. Multi-model combinations.

Example implementation:

Given the diversity of human disease progression, we anticipated a significant amount of data mismatch, suggesting that the learning algorithm should be robust against conditions not encountered during training. We also anticipate that the problem is nonlinear, as in most real-world problems.

After carefully considering a number of learning algorithms, we determined a combination of multivariate linear regression and base normalization network (BRN), and found consensus between the two. Linear regression was chosen due to its robustness in the presence of data mismatch. Non-parametric and boundary-determining learning algorithms based on L-2 objective functions are used instead of wild guesses (i.e., interpolation in engineering) when forced to operate outside their comfort zone. It is known to produce extrapolation.

Despite its robustness to data mismatch, linear regression is performed out of breath when modeling complex and nonlinear relationships. To eliminate this drawback, we combine base normalization network (BRN) and linear regression. In addition, leveraging the concepts of normalization and BIC, introducing additional penalty terms as a function of model complexity, thereby predicting the need for robustness for new and invisible data through generalization. Trade off needs to minimize errors.

The robustness of learning can be identified by noting the performance difference between random cross validation and full-data-set training, which is generally desirable in real-time implementations because they want to model as much real data as possible. 3A-3H show two selected learning algorithms with two different data sets in operation. 3A and 3B show linear regression for the first data set, and FIGS. 3C and 3D show the BRN for that data set. 3B and 3D are for the full data set, and FIGS. 3A and 3C are cross validation. 3A is a graph of actual (x axis) versus predicted (y axis) output with R _sq = 0.86358. 3B is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.83052. 3C is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.91497. 3D is a graph of the actual (x axis) vs. predicted (y axis) output with R _sq = 0.92077. 3E-3F show the BRN for that data set. 3F and 3H are for the full data set, and FIG. 3E is cross validation. 3E is a graph of actual (x axis) versus predicted (y axis) output with R _sq = 0.58165. 3F is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.56736. 3G is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.59826. 3H is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.59819. As expected, there is little performance discrepancy between random cross-validation training and full-data-set training on two different data sets, and even negative in some instances that are a testament to learning robustness. Do. 4A is a graph of actual (x-axis) vs predicted (y-axis) output with R _sq = 0.90001. 4B is a graph of actual (x axis) vs predicted (y axis) output with R _sq = 0.96846. In this case, the performance mismatch is rather pronounced, and if this situation is observed with robust learning algorithms, the level of data mismatch is very high.

If the learning algorithm simply remembers the solution instead of generalizing, we expect to show a greater performance mismatch than that shown in Figures 4A and 4B. 4A and 4B show the memory effect on multilayer perceptron. 4A is based on cross validation while FIG. 4B is based on full data set training using neural networks that minimize L-2 objective functions.

5 is a graph of multivariate linear regression (MVLR) prediction (x-axis) versus base normalization network (BRN) prediction (y-axis), from a line showing complete agreement, while the two learning algorithms generally agree in some cases. As can be seen by the numerous scattering points in the distance, they give different answers. Here, the predicted variable is the square root of the next one year cost (medical and medicament) using the previous 9 months of demographic data and claims.

In summary, we use the following criteria to ensure that our solutions are accurate and robust: minimizing errors; Minimize model-complexity; Model-weighted minimization for smooth, generalized responses; Feature ranking to find a point of diminishing return; Minimizing discrepancies between full-data-set training and 5- / 10-fold cross validation; And multi-model combinations.

As an example implementation of a hierarchical tree network with leaf-specific feature optimization, we will look at how some leaves behave differently.

For members with catastrophic conditions, very useful features arise from cost trends associated with chronic conditions such as renal disease / chronic kidney disease, cancer and rare diseases. Another interesting feature addresses recent hospital outpatient trends, where dialysis is routinely performed. In addition, the ratio of medical to Rx costs is important because we believe that higher Rx costs with correspondingly lower medical costs in this population segment are typically better health care.

Interestingly, those having both medical and Rx benefits, only having Rx claims using Humana, are not important, but the ratio of the cost of chronic Rx care divided by the total Rx cost and the amount of drug claim are both important. That is, depending on the type of benefit plan you have, the feature optimization algorithm automatically selects the subset optimally.

On the other hand, people with diabetes have mental disorders (those who have recently been diagnosed with a chronic condition tend to suffer from depression and other mental anxiety), vents, congestive heart failure, and endocrine. , Costly drivers such as digestive and other complications. This indicator is the number of different International Classifiction of Disease (ICD) codes over the last nine months. The ICD code is often referred to as the ICD9 code, and can also be seen as the 9 ^th revision of the code.

Nevertheless, there is no clear rule of thumb that can be used a priori to speculate that features will be more effective for a particular leaf. To investigate this problem, we have the following covariates as shown in FIGS. 6A-6F: number of unique drugs taken (FIG. 6A); Total Rx cost (FIG. 6B); Age (FIG. 6C); Transplant-related costs (FIG. 6D); Chronic cost (FIG. 6E); And multiple two-dimensional scatter plots of the square root of the cost of chronic artery disease (CAD) (FIG. 6F) versus the one-year future cost. If a simple linear relationship exists, a monotonic increasing line of straight lines will appear. Unfortunately, none of these features show this clear trend, indicating that there is no simple linear insight.

The first myth proved to be wrong is that the more medication you take, the more he or she will spend in the future. 6A is not shown for this relationship. In FIG. 6B, the overall Rx cost seems more likely, even if it represents a large portion (more than 90% of the population) in the initial half showing a non-than-conclusive picture. 6C shows that the most dangerous age group is about 50 to 65 years old. 6C may also indicate that age is not useful for predicting future costs. The same observation is valid for both the chronic cost in FIG. 6E and the CAD cost in FIG. 6F. In other words, having multiple chronic comorbid conditions does not automatically lead to unpleasant future outcomes. This observation aggregates the differences involved in generating predictive models for predicting future health conditions.

Therefore, instead of forcing our intuitions that are frequently wrong about predictive models or putting everything into the model (including a number of useless miscellaneous things), we construct a hierarchical tree network and allow the optimization algorithm to Classify and select the optimal subset for each leaf. In other words, let the data tell us what to implement for each leaf.

The present invention includes a computer-based system for predicting a future health condition of an individual: a. A hierarchical tree network assigning each member to one of the thirty nodes as a function of five stratification variables; b. Feature-subset selection for each node using a number of stochastic and combinatorial optimization algorithms; A separate learning and scoring module according to the operation mode; d. An arithmetic mean operator to find a match from two learning algorithms; And e. Report generator that outputs relevant parameters to aid in clinical intervention. Since five meaningful stratification variables are used to design the hierarchical tree network, the turnkey solution is passed in terms of data and life cycle requirements. In other words, we do not have to deal only with medical and Rx data and with more than six months of billing experience. In addition, the two learning algorithms are complementary to each other and therefore work in robust subspaces of vector space that include model-mismatch-to-data-mismatch and modeling-flexibility-to-model-complexity spectra.

The foregoing detailed description is primarily for providing clarity of understanding, and should not be understood as unnecessary limitations, and those skilled in the art will appreciate that various modifications may be made without departing from the spirit of the invention.

Claims (13)

In the method of predicting the future health condition of the individual, a. Establishing a hierarchical tree network that assigns each member only to at least one of the plurality of nodes as a function of a plurality of stratification variables for the plurality of members; b. Providing, to a computer-based system, member demographic data, active member medical billing data, and active member medical billing data for the plurality of members and each of the members; c. For each said node an optimal serve of features from a set comprising at least one of said member demographic data, said effective member medical claim data and said effective member medical claim data for all members assigned to said node; Performing feature selection on each of the plurality of nodes to identify a set; d. Parameters trained in the database to train an MVLR algorithm and a BRN algorithm using at least one of the member demographic data, the active member medical claim data and the effective member medical claim data, and generate a learned parameter database. Storing the; e. Utilize the learned parameter database, use member demographic data, the active member medical claim data and the effective member medical claim data of the one or more members for one or more of the members, and use the MVLR algorithm to MVLR future Calculating a health status score, calculating a BRN future health status score using the BRN algorithm, and determining a final score by calculating an arithmetic mean of the MVLR future health status score and the BRN future health status score. Characterized by the method for predicting a medical condition. The method of claim 1, After determining the final score, further comprising generating a member clinical condition report, wherein the member clinical condition report includes member identification information and the final score. A medical condition prediction method, characterized in that. The method of claim 2, Wherein said member clinical status report further comprises all or part of any of said member's medical billing information and medication billing information, and said member's clinical status data. The method of claim 1, In establishing the hierarchical tree network, the plurality of stratification variables may include a member's enrollment duration, presence of a member of a catastrophic condition; A health condition characterized by the presence or absence of a member of the diabetes flag, a member having a subscription trigger or an inpatient admission trigger, and at least one of the member claim data Forecast method. The method of claim 4, wherein If the member's registration period is less than six months, the member will be assigned to one of the first set of nodes; And if the registration period of the member is more than six months, the member is assigned to one of the second set of nodes. The method of claim 5, If the registration period of the member is less than six months and the member has a catastrophic state, the member will be assigned to the first subset of the first set of nodes; If the member's registration period is less than six months and the member has a absence of catastrophic status, the member is one of the first set of nodes that is not present in the first subset of the first set of nodes. One of the first set of nodes that will be assigned to and not present in the first subset of the first set of nodes comprises a second subset of the first set of nodes; If the member's registration period is at least six months and the member has a presence of a global state, the member will be assigned to a first subset of the second set of nodes; If the member's registration period is six months or more and the member has a absence of catastrophic status, the member is assigned to one of the second sets of nodes that are not present in the first subset of the second set of nodes. And wherein one of the second sets of nodes not present in the first subset of the second set of nodes comprises a second subset of the second set of nodes. . The method of claim 6, If the member's enrollment period is less than 6 months and the member has a absence of catastrophic status, the member has a diabetes flag and the member has a subscription trigger, the member is responsible for the first set of nodes. Assigned to a first subset of the second subset; If the member's enrollment period is less than six months and the member has a absence of catastrophic status, the member has a presence of a diabetic flag and the member has an inflation admission trigger, the member is responsible for the first set of nodes. Assigned to a second subset of the second subset; If the member's enrollment period is less than six months and the member has a catastrophic state, the member has a diabetic flag and the member has a subscription trigger, the member is responsible for the first set of nodes. Assigned to a third subset of the second subset; If the member's enrollment period is less than six months and the member has a catastrophic state, the member has a diabetic flag and the member has an inflation admission trigger, the member is responsible for the first set of nodes. Assigned to a fourth subset of the second subset; If the member's registration period is at least six months and the member has a catastrophic state, the member has a diabetic flag and the member has a subscription trigger, the member may be assigned to the second set of nodes. Assigned to a first subset of the second subset; If the member's enrollment period is at least six months and the member has a catastrophic state, the member has a diabetic flag and the member has an inflation admission trigger, the member may be assigned to the second set of nodes. Assigned to a second subset of the second subset; If the member's enrollment period is at least six months and the member has a catastrophic state, the member has a diabetic flag and the member has a subscription trigger, then the member is subject to the second set of nodes. Assigned to a third subset of the second subset; If the member's enrollment period is at least 6 months and the member has a catastrophic state, the member has a diabetes flag and the member has an inflation admission trigger, the member is responsible for the second set of nodes. And a fourth subset of the second subset. The method of claim 7, wherein The first subset of the first set of nodes, the first subset of the second subset of the first set of nodes, the second subset of the first set of nodes A second subset, the third subset of the second subset of the first set of nodes, the fourth subset of the second subset of the first set of nodes, the node The first subset of the second set of nodes, the first subset of the second subset of the second set of nodes, the second subset of the second subset of the second set of nodes A subset, the third subset of the second subset of the second set of nodes, and the fourth subset of the second subset of the second set of nodes, respectively, And a third node in which the member has medical and billing data, a second node in which the member has medical billing data only, and a third node in which the member has only medical billing data. The method of claim 4, wherein Said member will have a catastrophic condition if said member suffers from at least one of any cancer, terminal disease, transplantation, rare disease, HIV, coronary artery disease and a combination of chronic heart failure and hypertension State prediction method. The method of claim 1, The plurality of nodes of the hierarchical tree network comprises thirty nodes, a. A first node for any member having a registration period of less than six months, presence of catastrophic status, and both medical and medical billing data; b. A second node for any member having a registration period of less than six months, presence of catastrophic status, medical billing data only; c. A third node for any member having a registration period of less than six months, presence of catastrophic status, medical billing data only; d. A fourth node for any member having a enrollment period of less than six months, absence of catastrophic status, presence of diabetes triggers, presence of prescription triggers, and both medical and medical billing data; e. A fifth node for any member with a registration period of less than 6 months, absence of catastrophic status, presence of diabetes triggers, presence of prescription triggers, and medical billing data only; f. A sixth node for any member with a registration period of less than six months, absence of catastrophic status, presence of diabetes triggers, presence of subscription triggers, and medical billing data only; g. A seventh node for any member having a enrollment period of less than six months, the absence of catastrophic status, the presence of a diabetic trigger, the presence of an intake admission trigger, and both medical and medical billing data; h. An eighth node for any member having a enrollment period of less than six months, absence of catastrophic status, presence of diabetes triggers, presence of intake admission triggers, and medical billing data only; i. A ninth node for any member having a enrollment period of less than six months, the absence of catastrophic status, the presence of a diabetic trigger, the presence of an intake admission trigger, only medical billing data; j. A tenth node for any member having a enrollment period of less than six months, absence of catastrophic status, absence of diabetes triggers, presence of prescription triggers, both medication and medical billing data; k. An eleventh node for any member with a registration period of less than six months, absence of catastrophic status, absence of diabetes triggers, presence of prescription triggers, and medical billing data only; l. A twelfth node for any member having a registration period of less than 6 months, absence of catastrophic status, absence of diabetes triggers, presence of subscription triggers, medical billing data only; m. A thirteenth node for any member having a enrollment period of less than six months, the absence of catastrophic status, the absence of a diabetic trigger, the presence of an intake admission trigger, both medical and medical billing data; n. A fourteenth node for any member having only a enrollment period of less than six months, the absence of catastrophic status, the absence of a diabetic trigger, the presence of an intake admission trigger, and only medical billing data; o. A fifteenth node for any member having enrollment periods less than six months, absence of catastrophic status, absence of diabetes triggers, presence of inflation admission triggers, and medical billing data only; p. A sixteenth node for any member having a registration period of six months or more, presence of catastrophic status, both medical and medical billing data; q. A seventeenth node for any member having a registration period of six months or more, the presence of catastrophic status, and medical billing data only; r. An eighteenth node for any member having a registration period of six months or more, presence of catastrophic status, medical billing data only; s. A nineteenth node for any member having a enrollment period of six months or more, absence of catastrophic status, presence of diabetes triggers, presence of prescription triggers, and both medical and medical billing data; t. A twentieth node for any member having a enrollment period of six months or more, the absence of catastrophic status, the presence of a diabetic trigger, the presence of a subscription trigger, and medical billing data only; u. A twenty-first node for any member having a registration period of at least six months, absence of catastrophic status, presence of diabetes triggers, presence of subscription triggers, and medical billing data only; v. A twenty-second node for any member having a enrollment period of six months or more, the absence of catastrophic status, the presence of a diabetic trigger, the presence of an intake admission trigger, and both medical and medical billing data; w. A twenty-third node for any member having a enrollment period of six months or more, the absence of catastrophic status, the presence of a diabetic trigger, the presence of an intake admission trigger, and only medical billing data; x. A twenty-fourth node for any member having a enrollment period of six months or more, absence of catastrophic status, presence of diabetes triggers, presence of intake admission triggers, and medical billing data only; y. A twenty-fifth node for any member having a enrollment period of six months or more, absence of catastrophic conditions, absence of diabetes triggers, presence of prescription triggers, and both medical and medical billing data; z. A twenty-sixth node for any member having only a enrollment period of six months or more, the absence of catastrophic conditions, the absence of a diabetic trigger, the presence of a subscription trigger, and medical billing data only; aa. A twenty-seventh node for any member having only a enrollment period of six months or more, absence of catastrophic status, absence of diabetes triggers, presence of subscription triggers, and medical billing data only; bb. A twenty-eighth node for any member having a enrollment period of six months or more, the absence of catastrophic status, the absence of a diabetic trigger, the presence of an intake admission trigger, and both medical and medical billing data; cc. A twenty-ninth node for any member having only a enrollment period of six months or more, the absence of catastrophic status, the absence of a diabetic trigger, the presence of an intake admission trigger, and only medical billing data; And dd. And a thirtieth node for any member having only a enrollment period of six months or more, absence of catastrophic status, absence of diabetes triggers, presence of inflation admission triggers, and medical billing data only. The method of claim 10, After determining the final score, further comprising generating a member clinical status report, wherein the member clinical status report includes member identification information and the final score. The method of claim 11, Wherein said member clinical status report further comprises all or part of any of said member's medical billing information and medication billing information, and said member's clinical status data. The method of claim 10, Said member will have a catastrophic condition if said member suffers from at least one of any cancer, terminal disease, transplantation, rare disease, HIV, coronary artery disease and a combination of chronic heart failure and hypertension State prediction method.

Images