CN106446566A - Elderly cognitive function classification method based on random forest - Google Patents

Abstract

The invention relates to an elderly cognitive function classification method based on a random forest and belongs to the technical field of biomedicine. The method disclosed by the invention comprises the following steps: dividing the elderly cognitive function into three types by adopting MMSE scale scores and education levels; extracting a key cognitive domain influencing elderly cognitive function category classification by utilizing a cognitive function score relative ratio calculation method and a Pearson linearly dependent coefficient calculation method; establishing a random forest regression model, calculating attribute significance scores of non-scale attributes, and extracting external related attributes influencing the elderly cognitive function category classification; finally, equalizing the sample set based on the extracted key cognitive domain and external related attributes by adopting an SMOTE up-sampling method, and establishing an elderly cognitive function classification model by utilizing the random forest method. Compared with a scale classification method, the method provided by the invention has the advantages that the adopted attributes are few and easy to collect, and the method has high convenience; compared with other machine learning algorithms, subdivision of the elderly cognitive function types is realized, and research of a method for performing targeted intervention on the elderly cognitive functions is facilitated.

Description

Old people's cognitive function sorting technique based on random forest

Technical field

The present invention relates to a kind of old people's cognitive function sorting technique based on random forest, belongs to biomedical technology neck Domain.

Background technology

In recent years, China's aging population quantity grows with each passing day, and China has stepped into aging society.Senior health and fitness asks Topic becomes the significant concern object of society's medical treatment and security system.Human physiological functions can be continuous with the process of aging Decay, when physiological function decay to a certain extent when, old people just loses self care ability, now need a large amount of manpowers and Material resources ensure its life, can bring immense pressure to society's medical treatment and security system.Therefore how to extend senior health and fitness's life State, ensures life of elderly person self-care ability, extends human body physiological longevity with bigger social benefit than simple, is that population is old The key issue for facing is had to during age.And in senior health and fitness's animation, cognitive function has important ground Position, while affecting physiology and the mental health of this old people, is the important ring for ensureing senior health and fitness's animation.

Cognitive function refers to a series of Premium Features that human brain has, and such as perceives, understands, remembers and calculates Deng the i.e. general designation of the conscious ergasia of people, is the important component part for ensureing human body healthy living state.Cognitive function bag Multiple cognitive territories are included, has such as been calculated, structural capacity and language understanding etc., reflected the different aspect of brain function.Cognitive function Decay mainly occurs abnormal by the brain course of processing related to senior intelligent behaviors such as memory, thinking and study and causes, table The aspect such as elongated with the response time that present cognition speed slows down, memory ability goes down, has a strong impact on the orthobiosiss of people.Old people In age propagation process, cognitive function meeting slow-decay, this belongs to a kind of normal physiological phenomena, but this decay is with individuality Diversity, and can be affected by external environment and intervention to a certain extent, therefore realize dividing for old people's cognitive function quality Class, the prevention and early warning to the decay of old people's cognitive function is significant.

Research to old people's cognitive function sorting technique, employs the form of scale mostly at present.Scale is broadly divided into Two classes, a class be based on functional check mode, another kind of based on behavioral activity observation.In all of scale, simple essence Refreshing state Screening Scale (MMSE) is most widely used, and which is set up in 1975 by Folstein, belongs in functional check mode Based on scale, mainly include to orientation force (place and time orientation), memory (short-term and immediate memory), linguistic competence etc. The investigation of project.Scale total score is 30 points, the required duration about 5 to 10 minutes of filling in a form, and scale performance is good, and recall rate is permissible Reach 80%～90%.In addition draw clock test (CDT) and 7min neuro-cognitive Screening Scale falls within first kind scale.Wherein CDT tests structural capacity and the executive capability of experimenter by imitating picture clock and drawing the clock two ways of the time of specifying, and scale is total It is divided into 16 points.7min neuro-cognitive Screening Scale then includes to remember, orients, the test of language fluency and draw clock test totally 4 sides Face, it is stipulated that take as 7 minutes.The scale of observation human body behavioral activity is in the way of observing measured's behavioral activity, cognitive to which Function is classified, and conventional scale is mainly activities of daily living scale (ADL).ADL scale includes that body is taken care of oneself Scale and instrumental daily living measuring scale two parts, 14 altogether, the scale can (individual event be obtained by three kinds of different scores Point, PTS or subscale score) scale result is analyzed.

In recent years, the focus that old people's cognitive function disaggregated model is also research is built using machine learning method. Daoqiang Zhang is calculated using support vector machine (SVM) on the basis of tri- kinds of biomarkers of FDG-PET, MRI and CSF Method, constructs cognitive function disaggregated model, and the category of model accuracy rate reaches as high as 93.2%.2014, Gray was directed to low text Change horizontal crowd, factorial regression method and logistic regression method is combined, construct a cognitive function classification tool, the instrument Under Receiver operating curve, area (AUC) and recall rate are respectively 0.871 and 91.7%.

In sum, the form of scale, same opportunity, in the research of old people's cognitive function sorting technique, are employed mostly Device learning algorithm also begins to gradually be employed.Sorting technique with scale as instrument, exports score or grade to old according to scale Year people's cognitive function is classified, and with higher accuracy rate and credibility, but Scale and questionnaire project is many, using loaded down with trivial details, no Method meets the demand quickly divided by old people's cognitive function classification.The old people's cognitive function for being built using machine learning algorithm Disaggregated model, mostly two classification model (as realized the classification good and poor to cognitive function), draw to cognitive function classification Divide not enough finely, be unfavorable for realizing the research to individual targeted interference method.

Content of the invention

Present invention aim to address there is not convenient, offending problem in old people's cognitive function sorting technique, carry Go out a kind of old people's cognitive function sorting technique based on random forest.

The design principle of the present invention is：Realized to old people's cognitive function class based on MMSE scale score and schooling Other division, score relative ratio of the comprehensive Different Cognitive functional category under each cognitive territory, and cognitive function classification with The linearly dependent coefficient in Different Cognitive domain, determines and the closely related cognitive territory of cognitive function classification, i.e. crucial cognitive territory.Then Cognitive function score regression model is built using random forests algorithm, is extracted and old people's cognitive function is classified with material impact Non- scale attribute, i.e., outer attribute.Based on crucial cognitive territory and attribute outward, structure random forest disaggregated model, realize Classification to old people's cognitive function.The invention utilizes easy acquisition index, it is possible to achieve old people's cognitive function good, general and The other subdivision of poor three species.

The technical scheme is that and be achieved by the steps of：

Step 1, analysis MMSE scale score and schooling attribute, divide old people's cognitive function classification, utilize Division of the MMSE scale to cognitive territory, statistics Different Cognitive functional category is calculated each and recognizes in the score average of each cognitive territory Know the score relative ratio of Different Cognitive functional category under domain, while calculating the line of Different Cognitive functional category and each cognitive territory Property correlation coefficient, comprehensive relative ratio and linearly dependent coefficient, extract crucial cognitive territory, concrete methods of realizing is：

Step 1.1, using MMSE scale scoring event and schooling attribute, it is good to divide old people's cognitive function Good, general and poor three kinds of classifications.

Step 1.2, according to scale attribute, counts score average of the Different Cognitive functional category in each cognitive territory, calculates Score relative ratio of the Different Cognitive functional category under Different Cognitive domain.Computational methods are：

Wherein, G_ijIt is the score average of j-th cognitive function classification in i-th cognitive territory, i span is 1 to 9 Integer, represents 9 cognitive territories respectively, and j value is 1,2 or 3, to represent cognitive function respectively good, general and poor.

Step 1.3, calculates the Pearson linearly dependent coefficient of cognitive function PTS and score under Different Cognitive domain, meter Calculation method is：

Wherein n represents that number of samples, Z and Y need to calculate the attribute of dependency for bidimensional.

Step 1.4,1.2 gained relative ratio of combining step and step 1.3 gained Pearson linearly dependent coefficient, determine The crucial cognitive territory closely related with cognitive function classification.

Step 2, using non-scale attribute as independent variable, builds cognitive function score using random forests algorithm and returns mould Type, calculates importance score of the non-scale attribute to cognitive function, synthesized attribute importance score sequence height and collection difficulty Size chooses the outer attribute of impact old people cognitive function classification.Concrete methods of realizing is：

Step 2.1, non-scale attribute needed for choosing and cognitive function PTS attribute build data set, according to Yu Xianxuanding (number ntree of tree and Split Attribute number m), build random forest regression model to optimized parameter, to producing in model construction I-th regression tree, calculate its mean square sesidual MSE with the outer data of respective bag_i, i≤ntree, therefore calculating tries to achieve ntree altogether Mean square sesidual.

Step 2.2, in the outer data of bag, adds noise at random to j-th non-scale attribute, builds the outer data of new bag, The mean square sesidual MSE of per regression tree is recalculated with the outer data of newly-built bag_ij', gained mean square sesidual is calculated in conjunction with step 2.1, The importance score of the attribute is calculated, computational methods are：

Wherein F_jFor the importance score of j-th attribute, ntree is created the number of tree by random forest.

Step 2.3, repeat step 2.2, calculate the importance score of all non-scale attributes.

Step 2.4, based on the Importance of Attributes score of step 2.3 gained, carries out Importance of Attributes sequence from high to low, comprehensive Close consideration Importance of Attributes and collection difficulty selection is outer attribute.

Step 3, based on crucial cognitive territory and outer attribute, using the less class of SMOTE top sampling method balance sample number Not, on the basis of the data set after balance, random forest disaggregated model is built, the subdivision of old people's cognitive function classification is realized, Concrete methods of realizing is：

Step 3.1, the crucial cognitive territory that is extracted according to step 1 and step 2 and outer attribute, rebuild data set S.

Step 3.2, based on data set S, using SMOTE top sampling method, in the less classification of sample size, according to institute Multiplying power K need to be up-sampled, and K closest sample is chosen to each original sample, then most adjacent with each of which in original sample respectively The new samples of radom insertion manual creation between nearly sample, obtain the data set S' after sample balance, the computational methods of new samples For：

N_ij=O_i+rand(0,1)*R(O_i,O_ij)

Wherein, N_ijFor the individual new samples of jth (j≤K), O_iFor original sample, O_ijFor O_iJ-th closest sample, rand (0,1) represent and produce a random number more than 0 less than 1, R (O_i,O_ij) represent original sample O_iTo its closest sample O_ij Distance.

Step 3.3, based on data set S', obtains optimized parameter needed for model construction using ten folding cross validation methods, adopts Old people's cognitive function disaggregated model is built with random forests algorithm.

Beneficial effect

Compared to scale sorting technique, what the present invention was created is adopted based on old people's cognitive function disaggregated model of random forest The property set for being combined with non-scale attribute and scale attribute, these attributes are easy to collection, simplify the use of model, favorably In the quick division of old people's cognitive function classification, old people's cognitive function self-appraisal can be preferably applied for.

Compared to machine learning methods such as support vector machine, the old people's cognition work(based on random forest that the present invention builds Energy disaggregated model achieves the other division of good, general and poor three species of cognitive function, and classification is finer, and it is right to be advantageously implemented The research of the targeted interference method of old people's cognitive function.

Description of the drawings

Fig. 1 is old people's cognitive function sorting technique schematic diagram proposed by the present invention

Fig. 2 is model Contrast on effect experimental principle figure in specific embodiment

Fig. 3 is that specific embodiment middle-aged and elderly people cognitive function sorting technique tests design sketch

Specific embodiment

In order to better illustrate objects and advantages of the present invention, the reality to the inventive method with reference to the accompanying drawings and examples The mode of applying is described in further details.

Test data comes from 13 and is located at 7 different hospitals of provinces and cities survey data of 2011～2012 years, and data have altogether 9503, tie up per data 482,9 broad aspect such as medical conditions, personal essential information, cognitive function are included, data sample is equal For old people of the age more than or equal to 60 years old.

Test process mainly includes four processes, and all links are all completed on same computer, the allocation of computer For:Intel double-core CPU (dominant frequency 2.93GHz), 4GB internal memory, windows7 operating system.

Link one

This link describes the extraction of the crucial cognitive territory of impact old people's cognitive function category division in detail.It is embodied as step Rapid as follows：

Step 1.1, using MMSE scale score and schooling attribute, divides old people's cognitive function classification.

Old people's cognitive function category division criterion is shown in Table 1：

1. old people's cognitive function category division criterion of table

Step 1.2, extracts 9 and ties up cognitive territory score attribute and class label, after carrying out data scrubbing, obtains and includes 4516 The new data set of 10 dimension attributes.

Step 1.3, calculates score average of the Different Cognitive functional category in each cognitive territory, according to score relative ratio meter Formula is calculated, obtains score relative ratio of the Different Cognitive functional category under Different Cognitive domain.

Under Different Cognitive domain, each cognitive function category score relative ratio is shown in Table 2：

Each cognitive function category score relative ratio under 2. Different Cognitive domain of table

Step 1.4, by 9 dimension cognitive territory score attribute read group total cognitive function PTSs, according to cognitive function PTS With score under Different Cognitive domain, Pearson linearly dependent coefficient between the two is calculated.

Cognitive function and cognitive territory linear dependence analysis in table 3：

3. cognitive function of table is analyzed with cognitive territory linear dependence

Step 1.5, is typically less than 0.9 with good score relative ratio with cognitive function, and cognitive function is poor to be obtained with good Split-phase reduced value is less than 0.7, and Pearson correlation coefficient absolute value chooses composition ability, note meter collection more than 0.45 for standard And three cognitive territories of impermanent memory are used as the crucial cognitive territory of impact old people's cognitive function category division.

Link two

This link is described in detail realizes the outer of impact old people's cognitive function category division using random forest regression model Attribute extraction process.Specific implementation step is as follows：

Step 2.1, extracts non-scale attribute and cognitive function score attribute from original data source, screens out deficiency of data And noise data, obtain the new data set comprising 3477 103 dimension attributes.

Step 2.2, based on new data set, must be divided into dependent variable with cognitive function, build random forest regression model, adjust Regression tree number ntree and Split Attribute number m in model construction process, mean square using ten folding cross validation method computation models Residual values, preference pattern mean square sesidual value least model corresponds to parameter (ntree=480, m=21) as optimized parameter.

Step 2.3, builds random forest regression model based on model optimized parameter, per recurrence that model construction is produced Tree, calculates mean square sesidual MSE with the outer data of its respective bag_i, therefore produce altogether 480 mean square sesidual.

Step 2.4, in the outer data of bag, adds noise to each attribute respectively, builds the outer data of new bag, using structure The outer data test gained mean square sesidual MSE of back pkt._ij', mean square sesidual MSE original with step 2.3 gained_i, according to respective formula meter Calculate the importance score of each attribute.

Step 2.5, based on the Importance of Attributes score of step 2.4 gained, carries out Importance of Attributes sequence from high to low, comprehensive Consideration Importance of Attributes and collection difficulty is closed, schooling, date of birth, economic level is chosen, ridden public transportation means, live Environment, body-mass index and financing ability, 7 dimension attributes are used as the outer affiliation of impact old people's cognitive function category division altogether Property.

7 dimensions outreach Importance of Attributes score and are shown in Table 4：

The dimension of table 4.7 outreaches Importance of Attributes score

Link three

This link describes the building process of the old people's cognitive function disaggregated model based on random forest in detail.It is embodied as Step is as follows：

Step 3.1, according to the crucial cognitive territory that extracts and outer attribute 10 dimension attributes altogether, and cognitive function classification category Property, rebuild new data set.

Step 3.2, using SMOTE top sampling method, concentrates insertion manual construction sample in legacy data, by former inhomogeneity Very this quantitative proportion is from 18:4.65:1, balanced is 2:2:3.

Step 3.3, based on the data set after sample equilibrium, obtains disaggregated model using ten folding cross validation methods and builds institute Optimized parameter (regression tree number ntree=14 and Split Attribute number m=8) is needed, and then is built using random forests algorithm old Year people's cognitive function disaggregated model.

Link four

This link is described in detail by contrast experiment, by combining crucial cognitive territory and outer attribute, and is adopted on SMOTE Quadrat method, the lifting to old people's cognitive function disaggregated model effect.Specific implementation step is as follows：

Step 4.1, extract respectively 7 outer attribute, 3 crucial cognitive territories and both combine under 10 dimension attributes build 3 Individual data set S₁、S₂And S₃.

Step 4.2, based on 3 data sets of step 4.1 gained, is respectively adopted SMOTE top sampling method equalization sample, structure Build 3 new data set S₁'、S₂' and S₃'.

Step 4.3, balanced for sample 6 data sets altogether build disaggregated model respectively, calculate each model to not With the recall rate of cognitive function classification, and contrasted.

Modelling effect contrast is shown in Table 5：

5. modelling effect of table is contrasted

As can be seen from Table 4, structure old people's cognitive function disaggregated model integration test best results of the present invention, explanation Sample equalization processing and attribute fusion improve category of model performance jointly.

The present invention proposes one for there is not convenient, offending problem in current old people's cognitive function sorting technique Plant the old people's cognitive function sorting technique based on random forest.Proved by old people's cognitive function classification experiments, the present invention Higher recall rate is respectively provided with to Different Cognitive functional category, needed for model, attribute number is few, it is easy to gather, is easy to old people to recognize Know the quick division of functional category, while model carries out the other subdivision of three species to old people's cognitive function, it is right to be advantageously implemented The research of the targeted interference method of old individual human.

Claims (4)

1. a kind of old people's cognitive function sorting technique based on random forest, it is characterised in that methods described includes following step Suddenly：

Step 1, comprehensive MMSE scale score and schooling attribute, divide old people's cognitive function classification；

Step 2, divides gained cognitive function classification based on step 1, calculates Different Cognitive functional category obtaining in each cognitive territory Split-phase reduced value, calculates the Pearson linearly dependent coefficient of cognitive function and Different Cognitive domain, comprehensive score relative ratio and line Property correlation coefficient, extract impact old people's cognitive function category division crucial cognitive territory；

Step 3, using non-scale attribute as independent variable, builds cognitive function score regression model, meter using random forests algorithm Importance score of the non-scale attribute to cognitive function, synthesized attribute importance score and collection difficulty is calculated, chooses impact old People's cognitive function category division outer attribute；

Step 4, based on crucial cognitive territory and outer attribute, structure random forest disaggregated model, realizes old people's cognitive function class Other division.

2. a kind of old people's cognitive function sorting technique based on random forest according to claim 1, it is characterised in that The extraction of crucial cognitive territory, concrete methods of realizing is：

Step 2.1, according to scale attribute, counts score average of the Different Cognitive functional category in each cognitive territory, calculates different Score relative ratio of the cognitive function classification under Different Cognitive domain, computational methods are：

${\mathrm{GR}}_{ij}=\frac{G_{ij}}{G_{i1}}$

Wherein, G_ijBe the score average of j-th cognitive function classification in i-th cognitive territory, i span be 1 to 9 whole Number, represents 9 cognitive territories respectively, and j value is 1,2 or 3, to represent cognitive function respectively good, general and poor；

Step 2.2, calculates the Pearson linearly dependent coefficient of cognitive function PTS and score under Different Cognitive domain, calculating side Method is：

${\mathrm{\ρ}}_{Z,Y}=\frac{\mathrm{\Σ}ZY-\frac{\mathrm{\Σ}Z\mathrm{\Σ}Y}{n}}{\sqrt{({\mathrm{\ΣZ}}^2-\frac{{\left(\mathrm{\Σ}Z\right)}^2}{n})({\mathrm{\ΣY}}^2-\frac{{\left(\mathrm{\Σ}Y\right)}^2}{n})}}$

Wherein n represents that number of samples, Z and Y need to calculate the attribute of dependency for bidimensional；

Step 2.3, chooses score relative ratio and changes greatly, and the big cognitive territory of Pearson linearly dependent coefficient, as with cognition Functional category divides closely related crucial cognitive territory.

3. a kind of old people's cognitive function sorting technique based on random forest according to claim 1, it is characterised in that Attribute extraction outward, concrete methods of realizing is：

Step 3.1, creates new data set using non-scale attribute and cognitive function PTS attribute structure, and using adjusting, ginseng gained is optimum (number ntree of tree and Split Attribute number m) build random forest regression model to parameter, to per produced in model construction Regression tree, calculates mean square sesidual MSE with the outer data of its respective bag_i, i≤ntree, thus altogether calculating to try to achieve ntree mean square residual Difference；

Step 3.2, in the outer data of bag, adds noise at random respectively to each attribute, builds the outer data of new bag, using structure The outer data test gained mean square sesidual MSE of back pkt._ij', and the original mean square sesidual MSE of step 2.1 gained_i, computation attribute is important Property score, computational methods are：

$F_j=\frac{\underset{i=1}{\overset{n}{\Σ}}({\mathrm{MSE}}_i-{{\mathrm{MSE}}_{ij}}^{\′})}{ntree}$

Wherein F_jFor the importance score of j-th attribute, ntree is the number of random forest establishment tree；

Step 3.3, based on the Importance of Attributes score of step 3.2 gained, carries out Importance of Attributes sequence from high to low, comprehensively examines Consider Importance of Attributes and collection difficulty, choose the outer attribute of impact old people's cognitive function category division.

4. a kind of old people's cognitive function sorting technique based on random forest according to claim 1, it is characterised in that Old people's cognitive function disaggregated model builds, and concrete methods of realizing is：

Step 4.1, according to the crucial cognitive territory and outer attribute that extracts, rebuilds data set S；

Step 4.2, based on new data set S, adopts SMOTE top sampling method to minority class sample, according to required up-sampling multiplying power K, chooses K closest sample to each original sample, then respectively between the closest sample of original sample and each of which with The machine transplanting of rice enters the new samples of manual creation, obtains the data set S' after sample balance, and the computational methods of new samples are：

N_ij=O_i+rand(0,1)*R(O_i,O_ij)

Wherein, N_ijFor the individual new samples of jth (j≤K), O_iFor original sample, O_ijFor O_iJ-th closest sample, rand (0,1) Represent and produce a random number more than 0 less than 1, R (O_i,O_ij) represent original sample O_iTo its closest sample O_ijDistance；

Step 4.3, based on data set S' after equilibrium, obtains optimized parameter needed for model construction using ten folding cross validation methods, Old people's cognitive function disaggregated model is built using random forests algorithm.