Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 3, No. 1, 2023 

 

30 

Intelligent Diagnosis of Alzheimer's disease 

Ruifan Huang1, a, Huifei Xu2, b 

1 School of Sun Yat-sen University, Guangzhou 510000, China 
2 School of Hangzhou Normal University, Hangzhou 310000, China. 

a1059577592@qq.com, b1225604713@qq.com 

 

Abstract: In this paper, we combine the knowledge of decision tree model and K-means clustering algorithm with knowledge 

of statistics to solve the problems of Diagnosis of Alzheimer's disease using brain structure. In the step of selecting features, this 

paper considers both data-based and theoretical perspectives for the analysis of variance in Alzheimer's disease. Based on the 

data perspective, the value of variance represents the magnitude of data volatility, i.e., data with data volatility has more abilities 

to cause data changes in feature classification. Therefore, in this paper, a total of 28 indicators of Site-faq were selected as 

necessary indicators and histograms were made to observe the volatility. After the above data processing, we Use the attached 

structural brain features and cognitive behavioral features to design an intelligent diagnosis of Alzheimer's disease. For multi-

classification models, this paper uses the K-means algorithm to cluster and continue to cluster MCI into SMC, EMCI, LMCI to 

observe the difference between cluster centers. After clustering 53 indicators, this paper analyzes the three types of cluster centers, 

mainly to analyze the different points in the number of clusters, and analyzes the difference in the mean value of the 

characteristics of the cluster centers, and at the same time, this paper observes the differences between the characteristics by 

visualizing the data. In the end, we analyze them in relation to the time points to uncover patterns in the evolution of different 

categories of diseases over time. At the same time, due to the large amount of data, this paper only selects a small sample of five 

features for analysis. 

Keywords: Decision tree; Random forest; K-means clustering; Pearson coefficient. 

 

1. Introduction 

1.1. Problem Backgroud 

Alzheimer's disease (AD) is a progressive 

neurodegenerative disease with an insidious onset. 

It is characterized clinically by a full spectrum of dementia, 

including memory impairment, aphasia, dysfluency, agnosia, 

impairment of visuospatial skills, executive dysfunction, and 

personality and behavioral changes, the cause of which is still 

unknown. It is characterized by a progressive decline in the 

ability to perform activities of daily living, with various 

neuropsychiatric symptoms and behavioral disturbances. The 

disease is usually progressive in the elderly, with progressive 

loss of independent living skills and death from complications 

10 to 20 years after the onset of the disease. The preclinical 

stage of Alzheimer's disease, also known as mild cognitive 

impairment (MCI), is a transitional state between normal and 

severe. Due to the limited cognition of the disease by patients 

and their families, 67% of patients were diagnosed as 

moderate to severe and had missed the best intervention stage. 

Therefore, early and accurate diagnosis of Alzheimer's 

disease and mild cognitive impairment is of great significance 

[1] 

1.2. Clarifications and Restatements 

This paper uses the brain structural characteristics and 

cognitive behavioral characteristics of the different categories 

of people provided in the data set to construct an Alzheimer's 

disease identification model and design an intelligent 

diagnostic method to accurately diagnose Alzheimer's disease. 

In this paper, the research is decomposed into four problems 

shown below. 

1). Preprocess the characteristic indicators of the attached 

data to investigate the correlation between data characteristics 

and the diagnosis of Alzheimer's disease. 

2). Use the attached structural brain features and cognitive 

behavioral features to design an intelligent diagnosis of 

Alzheimer's disease. 

3). First, cluster CN, MCI and AD into three major classes. 

Then, for the three subclasses contained in MCI (SMC, EMCI, 

and LMCI), the clustering was continued to be refined into 

three subclasses. 

4). The same sample in the annex contains features 

collected at different time points, we will analyze them in 

relation to the time points to uncover patterns in the evolution 

of different categories of diseases over time. 

2. General Assumptions and Variable 
Description 

2.1. Assumptions 

1). Assume that the effect of environmental factors on the 

composition of ancient glass artifacts will not change the 

original chemical composition content ratio by 20%. 

2). Assume that in the sampling test, there is no human 

error in the test results, while the error between the test results 

and the true content should be less than 2%. 

3). It is assumed that chemical components not collected in 

the sampling tests and not studied will not have a significant 

correlative effect on the content of other chemical 

components. 

4). Assume that there are no interactions between different 

chemical components or between different representations of 

the artifacts. 

5). For taking into account the impact of ambient 

temperature and humidity on the test, assuming that the 

sampling process are in the same environment and the 

sampling time interval is short. 



 

31 

2.2. Variable Description  

Table 1. Symbol specification table 

Symbols Description 

 

The specific value of the   sample of 

corresponding to the   indicator of the 

parent series 

 
Subseries i.e. interpolation rate at the  

th sample  

 
Coefficient of influence of the  

variable on the classification 

 The  variable 

 
denotes the probability that the   th 

variable is classified as a barium lead 

variable 

 
Normalized value of a chemical 

component in the i-th data set 

 Cluster i 

 The sample points in the  

 Sample Means 

 
Distance of the point to all other points in 

the cluster it belongs to 

 
The average distance from a point to all 

points in a cluster that does not contain it 

 Contour coefficient of the point 

 
The rank difference, i.e. the position of the 

number after sorting two sets of data from 

smallest to largest 

3. Problem 1 

The process of modeling and solving Problem 1 is shown 

in Figure 1: 

 

Figure 1. Problem 1 solution process 

3.1. Task 1: Analysis of relationship 

3.1.1. Data pre-processing 

The flow chart for data pre-processing is as follows: 

 

Figure 2. Basic data pre-processing process 

For the codified data, this paper quantifies them according 

to the coding serial number, and for the more important codes, 

i.e., the text coding part, this paper chooses to encode them 

manually, for example, for the severity of AD, named as 

follows. 

 

Figure 3. Quantitative processing 

3.1.2. Model buiding and solving 

After the selection of indicators, in order to further 

determine the reasonableness of the selection of indicators. In 

this paper, ANOVA will be used to determine whether there is 

a significant difference in the change of each statistic over 

time. First, ANOVA is introduced in this paper as follows. 

The statistical inference method used in the one-way 

ANOVA is to calculate the F statistic and perform the F test. 

The total sum of squares of variance, denoted as SST, is 

decomposed into two parts: one part is the variance caused by 

the control variables, denoted as SSA (sum of squares of 

variance between groups); the other part is the SSE (sum of 

squares of variance within groups) caused by the random 

variables. The relationship between the two variables is 

shown below [2]. 

SST = SSA + SS 

In general, assume that there are k levels and establish the 

following assumptions. 

H0: u1 = u2=. . . uk 

cH1: u1, u2, . . . , uk, not all equal to each other 

The error within each overall system, mainly the random 

error caused by random sampling, is called the intra-group 

error, and the sum of squared intra-group errors is denoted as 

SSA. 

SSA = ∑ ni(xi̅ − x̅)2

k

i=1

 

The errors between the totals, mainly random errors and 

systematic errors caused by random sampling, are called 

inter-group errors, and the sum of squared inter-group errors 

is SSE. 

SSE = ∑ ∑(xij − xi̅)
2

ni

k

k

i=1

 

The total sum of squared errors of the data is denoted as 

SST. 

 

SST = ∑ ∑(xij − x̅)2

ni

k

k

i=1

 

Where denotes the jtℎ observation at the itℎ level, denotes 

the mean value at the jtℎ level, and denotes the mean value 

of the n observations. 

SST = ∑ ∑(xij − x̅)2 = SSA + SSE

ni

k

k

i=1

 

The above calculation shows that the total error of the data 

comes from two main sources: intra-group error on the one 

hand and inter-group error on the other. It can be shown that 

when the means of the k overall groups are equal, 
SSA

k − 1⁄ SSE
n − k⁄⁄  obeys the F distribution with degrees 

of freedom k-1 in the numerator and n - k in the denominator. 

F =
SSA (n − 1)⁄

SSE (n − k)⁄
≤ Fα(k − 1, n − k) 



 

32 

Each F value will correspond to a p value, and the larger 

the F value and the smaller the p value, the less likely it is to 

accept the original assumption, which is considered to be a 

large difference in characteristics between groups. Therefore, 

the smaller the p-value, the more this feature should be 

preserved.  

 

Figure 4. Analysis of variance histogram 

Pearson correlation analysis is a method used to study the 

closeness of association between two or more variables. In 

order to explore which factors are associated with the cause 

of Alzheimer's disease, so the correlation analysis algorithm 

is used, and the Pearson correlation value can be a good 

representation of the correlation between two variables. Some 

of the indicators in the experiment have an interaction 

relationship, and the Pearson correlation value is chosen to 

calculate the degree of correlation between different human 

physical characteristics affecting Alzheimer's disease, and the 

Pearson correlation value expression is shown below [3]. 

f(x, y) =
∑ (Xa − X̅)(Ya − Y̅)t

a=1

√∑ (Xa − X̅)2t
a=1 √∑ (Ya − Y̅)2t

a=1

 

In the equation, f(x, y)represents the similarity between the 

two variables X and Y, a represents the atℎ  sample of a 

certain indicator, t is the total number of samples, and X̅ and 

Y̅ represent the values of two variables, X and Y, respectively. 

Pearson correlation values f(x, y) always fall between -1 and 

1. When 0.8<a<1, the two variables are extremely strongly 

correlated; When 0.6<a<0.8, the two variables are strongly 

correlated; When 0.4<a<0.6, the two variables are moderately 

strongly correlated; When 0.2<a<0.4, the two variables are 

weakly correlated; When 0<a<0.2, the two variables are very 

weakly or not correlated. 

 

Figure 5. Correlation value line graph 

Therefore, in this paper, a total of 28 indicators of Site-faq 

were selected as essential indicators, and histograms were 

made to observe the volatility, and Pearson correlation 

analysis was done for the remaining variables with high 

volatility to screen the indicators with strong correlation to the 

generation and deterioration of Alzheimer's disease. Based on 

this, a total of 53 indicators were selected as basic indicators 

in this paper. 

4. Problem 2 

Problem 2 model building and solution process is shown in 

Figure 6; 

 

Figure 6. Problem 2 solution process 

4.1. Model buiding and solving 

Decision trees are supervised learning, a typical tree model, 

which embodies a functional mapping between features and 

labels. The non-leaf nodes represent the division of the 

sample on a feature, and the sample set is divided into several 

sub-trees according to the different values of the decision tree 

is divided into several sub-trees according to the different 

values of the feature, until the leaf node or the condition is 

satisfied. or a condition is met, the sample set is not split again. 

corresponds to a classification. The core problem of 

constructing a decision tree is how to select a suitable feature 

to split the current sample set. 

 

Figure 7. Decision tree structure diagram 

There are two types of decision trees, depending on the 

target variable. 

(1) Categorical variable decision tree: The target variable 

is classified according to the input characteristics. For 

example, suppose you are asked to predict the relative price 

of a computer into one of the following three categories: low, 

medium, or high. Features may include CPU, memory size, 

hard disk size, graphics card, sound card, etc. The decision 

tree will learn from these features, and after passing each data 

point to each node, it will eventually reach the 3 classification 

goals: low, medium, and high. 

(2) Continuous variable decision tree: In this case, the 

features input to the decision tree (e.g. quality of a house) will 

be used to predict the continuous output (e.g. price of a house). 



 

33 

 

Figure 8. SHAP value (impact on model output) 

1). Build a gradient boosted tree (GBDT) classification 

model from the training set data. 

2). Calculate the feature importance from the established 

Gradient Boosted Tree (GBDT). 

3). Apply the established Gradient Boosted Tree (GBDT) 

classification model to the training and testing data to obtain 

the classification evaluation results of the model. 

4). Since the gradient boosting tree (GBDT) is random, the 

result of each operation is not the same, if this training model 

is saved, the subsequent data can be directly uploaded to this 

training model to calculate the classification. 

5). Note: Gradient Boosting Tree (GBDT) cannot get the 

definite equation like the traditional model, and the model is 

usually evaluated by testing the data classification effect. 

During the development of machine learning models, it is 

expected that the trained model will perform well on new, 

unseen data. To simulate new, unseen data, data partitioning 

is performed on the available data, thus splitting it into 2 parts 

(sometimes called training-test partitioning). In particular, the 

first part is a larger subset of the data used as the training set 

(e.g., 80% of the original data) and the second part is usually 

a smaller subset used as the test set (the remaining 20% of the 

data). It is important to note that this data splitting is 

performed only once. Next, a prediction model is built using 

the training set, and this trained model is then applied to the 

test set (i.e., as new, unseen data) for prediction. The best 

model is selected based on its performance on the test set, and 

hyperparameter optimization is also performed in order to 

obtain the best model [4]. 

As the below figure is a mixed matrix, firstly, all samples 

are divided into positive samples P (labeled as positive) and 

negative samples N (labeled as negative), the positive 

samples that are still positive after model prediction are noted 

as True Positive (TP), the positive samples that are predicted 

to be negative are called False Positive (FP), similarly, the 

negative samples that are predicted to be negative are called 

True negative (TN), and the negative samples that are 

predicted to be positive are called False negative (FN). 

samples that are predicted to be negative are called True 

negative (TN) and negative samples that are predicted to be 

positive are called False negative (FN), so after model 

prediction there are P = TP+FN and N = FP+TN. 

 

Figure 9. Confusion Matrix Heat Map 

 

Figure 10. Feature Importance 

The below table shows the prediction evaluation metrics of 

the cross-validation set, training set and test set to measure the 

prediction effectiveness of GBDT through quantitative 

metrics. Among them, the evaluation metrics of the cross-

validation set can continuously adjust the hyperparameters to 

obtain a reliable and stable model. 

● Accuracy rate: the proportion of correctly predicted 

samples to the total samples, the larger the accuracy rate, the 

better. 

● Recall rate: The proportion of predicted positive samples 

out of the actual positive samples, and the larger the recall rate, 

the better. 

● Precision rate: The proportion of the predicted positive 

samples to the actual positive samples, and the larger the 

precision rate, the better. 

F1: The sum of precision and recall. Precision and recall 

affect each other, and although a high level of both is the 

desired ideal, in practice, it is often the case that a high 

precision rate results in a low recall rate, or a low recall rate 

results in a high precision rate. If a balance of both is needed, 

then the F1 metric can be used. 



 

34 

Table 2. Model evaluation results 

 

Accuracy 

Recall 

Accuracy 

Accuracy 

Recall 

Accuracy 

Accuracy 

Recall 

Accuracy 

F1 

Training 

set 
1 1 1 1 

Test set 0.689 0.689 0.687 0.687 

The random forest algorithm is a highly flexible machine 

learning algorithm proposed by Leo & Culte et al. in 2001, 

which is based on statistics and combines Bagging integrated 

learning theory and random subspace methods. The basic idea 

is to set many weak classifiers into one strong classifier. The 

random forest consists of several classification trees, each 

corresponding to a different training subset. Each training 

subset consists of multiple random samples of the training set 

with put-back by the Bootstrap sampling method.  

The k trees are constructed as described above, and these 

decision trees are grouped together to form a random forest. 

The random forest contains a second level of randomness, 

where m values are randomly selected from all the feature 

input values M to construct each decision tree; from these m 

values, each segmentation node is selected for optimization 

according to the principle of minimizing the Gini coefficient, 

thus reducing correlation and improving predictive power. 

The random forest simulation process does not perform the 

pruning process as in the decision tree. K predictors are 

obtained, and each predictor makes one prediction, and finally 

the predictions are aggregated and the average is calculated. 

The overall vector space is divided into smaller spaces by 

building individual decision trees, and the extracted decision 

trees are used as the training set.  

The set of samples that are not extracted is called out-of-

bag (OOB) data and is used as the validation set. The best 

branch is found by randomly selecting less than p alternative 

branches within each decision tree of the original data set. By 

adjusting the size of the leaf branches, the tree is stopped 

when it reaches 4 splits. The growth of the decision tree is 

controlled by adjusting the size of the leaf branches and 

stopping the formation of a single decision tree when it 

reaches 4 splits. A combinatorial model of the random forest 

algorithm is generated according to certain rules. 

 

Figure 11. Random forest model structure diagram 

When studying machine learning, we expect new, never-

before-seen material that has been trained to yield good 

results. To simulate the new, unseen material, the already 

existing material is divided into two blocks (sometimes called 

the training-testing area). Specifically, said first component is 

a subset of the larger data (e.g., 80% of the initial data), while 

said second component is typically a much smaller subset (the 

remaining 20% of the data). It is important to note that such a 

splitting of the data can be performed only once. Next, this 

training set is used to construct a forecast model and the 

model is used for the test set (i.e., new information that has 

not been seen). By examining its performance, an optimal 

model was selected and optimized in terms of 

hyperparameters, and a better mathematical model was 

obtained. 

A mixing matrix is shown below, where all samples are 

divided into P (marked as positive) and N (marked as 

negative), and samples that remain positive (TP) after 

prediction by the model are called false positive (FP) if they 

are predicted to be negative (FP), and if they are predicted to 

be negative (FP). The sample with negative prediction is 

called "TN", while the prediction for "FN" is "FN", so that 

after the model prediction, there will be P= TP+ FN, N= FP 

TP+ FN, N= FP+ TN. 

 

 

Figure 12. Feature Importance 

 

Figure 13. Feature Importance 

The below table shows the classification evaluation metrics 

for the training and test sets, and the quantitative metrics are 

used to measure the classification effectiveness of the 



 

35 

decision tree on the training and test data. 

● Accuracy: the proportion of correctly predicted samples 

to the total samples, the larger the accuracy, the better. 

● Recall rate: The proportion of predicted positive samples 

out of the actual positive samples, the larger the recall rate, 

the better. 

● Precision rate: The proportion of the predicted positive 

samples to the actual positive samples, and the larger the 

precision rate, the better. 

F1: The sum of precision and recall. Precision and recall 

affect each other, and although a high level of both is the 

desired ideal, in practice, it is often the case that a high 

precision rate results in a low recall rate, or a low recall rate 

results in a high precision rate. If both are needed, then the F1 

metric can be used. 

● oob_score: For classification problems, oob_score is the 

accuracy rate of out-of-bag data. If there is put-back sampling 

selected during the tree building process, about 1/3 of the 

records are not extracted. The ones that are not extracted 

naturally form a control data set that can be used for model 

validation. Therefore, the random forest does not need to set 

aside additional data for cross-validation; its own algorithm is 

similar to cross-validation, and the out-of-bag error is an 

unbiased estimate of the prediction error. 

Table 3. Model evaluation results 

 

Accuracy 

Recall 

Accuracy 

Accuracy 

Recall 

Accuracy 

Accuracy 

Recall 

Accuracy 

F1 

Training 

set 
0.842 0.842 0.852 0.844 

Test set 0.601 0.601 0.639 0.597 

Based on this, this paper uses decision tree model and 

random forest model to classify the data, and constructs 

model accuracy indexes: confusion matrix, acc, etc. so as to 

judge the advantages and disadvantages of the classification 

model. By comparing the judgment indexes of decision tree 

model and random forest model, it is found that decision tree 

model has better classification effect. 

5. Problem 3 

The problem 3 model building and solution process is 

shown in Figure 13. 

 

Figure 13. Problem 3 solution process 

5.1. Model Establishment and Solving 

The K-means clustering algorithm is a clustering algorithm 

that classifies based on a functional criterion, based on 

minimizing the clustering criterion function. The clustering 

criterion function is the sum of squares of the distances from 

each node in the cluster set to the center of the cluster of that 

class. For all M pattern classes criterion function is defined. 

K = ∑ ∑‖Xi − Zj‖

Nj

i=1

M

k=1

 

Where 𝑆𝑗 is the 𝑗𝑡ℎ cluster set with cluster center 𝑍𝑗; 𝑁𝑗 

is the number of samples contained in the jth cluster set Sj. 

The clustering criterion of KM: choose a cluster center Zk Let 

the criterion function K be extremely small, i.e., make the 

value of Kk extremely small, we get 

𝑍𝑘 =
1

𝑁
∑ 𝑋𝑖

𝑁𝑘

𝑖=1

 

Let there be N nodes, and the calculation steps are as 

follows: 

Any M initial cluster centers Z1(1), Z2(1), . . . ZM(1), M < 

N, and the serial numbers in parentheses represent the number 

of iterations. 

The remaining nodes are assigned to M cluster centers 

according to the shortest distance principle to one of the M 

cluster centers, i.e. 

min{‖X − Zi(m)‖, i = 1,2, . . . , M} = ‖X − Zk(m)‖ 

Where m is the order number of the iterative operation. 

Calculate the new vector values for each cluster center. 

𝑍𝑘(𝑚 + 1) =
1

𝑁𝑘
∑ 𝑋

𝑋∈𝑆𝑘(𝑚)

 

If Zk(m + 1) ≠  Zk ( m) , j =  1, 2, . . . , M,  then return to 

reclassify the nodes one by one and repeat the iterative 

computation, and vice versa, the algorithm converges and the 

computation is completed. 

Table 4. Model evaluation results 

 DX_bl SITE AGE 
PT_GEND

ER 

PT_EDU

CAT 

PT_ET

HCAT 

PT_RACC

AT 

PT_MAR

RY 

APO

E4 

1 106.52 70.86 2.00 16.26 1.18 1.34 1.64 0.40 0.13 

2 80.51 73.16 2.00 16.05 1.04 1.13 1.40 0.53 0.41 

3 81.93 73.41 2.00 16.20 1.04 1.13 1.39 0.48 0.04 

Therefore, in this paper, we first clustered the overall into 

three classes, denoted as MCI, CN\AD. Then, this paper used 

K-means algorithm for clustering, and continued to cluster 

MCI into SMC\EMCI\LMCI thereby observing the difference 

between the clustering centers. 

In this paper, we first clustered the overall into 3 classes, 

denoted as MCI, CN, and AD. Then, this paper used K-means 

algorithm for clustering and calculated the distribution of the 

values of 53 indicators under the three clustering centers. 

Under different clustering categories, the important factors 

affecting the classification are judged by comparing the 

magnitude of the values of the same indicators. 

The above visualization reveals that among the 53 

underlying indicators, a large number of indicators have little 

effect on clustering, which means that there is little difference 

between the values of the same indicators under different 

clustering categories, or even almost identical, but there are 

significant differences in variables such as FDG, AV45, FBB, 

and ABETA. Although the values differed little, their 

standardization revealed significant differences between the 

values of the same indicators under different clustering 

categories, which also provides a good basis for finding the 

etiology of Alzheimer's disease. 



 

36 

 
Figure 14. Visualization of clustering effects 

5.2. Model Establishment and Solving 

The clustering was continued by clustering the MCI into 

SMC, EMCI, and LMCI to observe the difference between 

the clustering centers. 

 

Figure 15. Visualization of clustering effects 

From the above visualization, it is clear that many of the 53 

basic indices have a small effect on the clusters, i.e.: in the 

classification, the differences are small and almost identical 

for the same index, but in FDG, AV45, FBB; different factors 

such as ABETA differ significantly. Despite the small 

difference between the two data, after normalization, we can 

see that there is a significant difference in the same index in 

each cluster under the same classification, which can find a 

good basis for the pathogenesis of Alzheimer's disease. 

 

Table 5. Model evaluation results 

 
DX_bl SITE AGE 

PT_GEND

ER 

PT_EDU

CAT 

PT_ETHC

AT 

PT_R

ACCAT 

PT_MA

RRY 

APO

E4 

SMC 107.04 70.1 2.00 16.15 1.09 1.37 1.73 0.35 0.12 

EMCI 79.75 72.3 2.00 16.08 1.04 1.14 1.38 0.56 0.41 

LMCI 80.97 72.5 2.00 16.15 1.05 1.15 1.37 0.54 0.04 

6. Problem 4 

The problem 4 model building and solution process is 

shown in Figure 16; 

 

Figure 16. Problem 4 solution process 

6.1. Model Establishment and Solving 

This paper first tests the model based on temporal 

indicators, which are the degree of change in the patient's 

condition from different times of testing, respectively. 

Therefore, this paper screens the tabular data so that it can be 

analyzed using statistical knowledge. Also, for further 

analysis, this paper argues that the evolutionary pattern of 

different categories of diseases over time can be seen as 

changes in brain characteristics and cognitive characteristics 

over the time of testing. Also, due to the large size of the data, 

only a small sample of the five characteristics is selected for 

analysis in this paper. 

 
Figure 17. Trend line graph of different brain features 



 

37 

 
Figure 18. Trend line graph of different brain features 

From the above discounted graph of data trends, it can be 

seen that most of the indicators have a tendency to increase, 

and the magnitude of the values of a few indicators maintains 

within a certain fluctuation range, which proves that along 

with the passage of time, there is a characteristic change 

process in the degree of disease symptoms, and it can be seen 

from the graph that the values of  RAVLT_immediate, 

RAVLT_learning, RAVLT_ forgetting and other indicators 

obviously have a decreasing trend, so when studying the 

changes of brain characteristics and cognitive characteristics 

with the detection time, this paper chooses to divide the data 

into behavioral characteristics and brain characteristics, and 

analyze different indicators under different types, where the 

data trend data of cognitive characteristics are shown below. 

 

 

 

Table 6. Trend data for different brain characteristics 

IMAGEUID Ventricles Hippocampus WholeBrain Entorhinal Fusiform MidTemp ICV 

118697 38712 4495 972134 2239 19818 20622 1524760 

118672 38374 0 964788 0 0 0 1513350 

35725 41869 0 971882 0 0 0 1514160 

59854 44002 0 931047 0 0 0 1527030 

83343 44139 0 945435 0 0 0 1523690 

129868 49536 0 952495 0 0 0 1504090 

205579 55599 0 912309 0 0 0 1524340 

7. Summary 

In this paper, we first determine the importance index of 

different indicators with Alzheimer's disease by constructing 

a correlation analysis through pre-processing of the data. We 

then use structural brain features and cognitive behavioural 

features to design an intelligent diagnostic model of 

Alzheimer's disease in which we choose to use a decision tree 

as a model for classification. 

Secondly, we used a time series model to analyse the 

relevant data by arguing that the evolutionary patterns of 

different categories of disease over time can be seen as 

changes in brain features and cognitive characteristics over 

the time of detection. 

Finally, we used the offering to review the etiology and 

onset symptoms of Alzheimer's symptoms, summarise and 

analyse the distribution of different indicators at different 

onset stages, and provide different early intervention and 

diagnostic criteria for the five categories of CN, SMC, EMCI, 

LMCI and AD. 

References 

[1] Du Jia, Fu Yan, Fan Jia, Zhou Miaorong & Zhang Yeting: 
Aerobic exercise intervention with Aβ_(1-42) induced changes 
in hippocampal synaptic structure and synaptic proteins in rats 
with Alzheimer's disease. Chinese Journal of Tissue 
Engineering Research. 

[2] Lv Romance & Li Aiping. (2022). Role of perineuronal matrix 
network in regulating cognitive function and Alzheimer's 
disease. Progress in Physiological Science (02),100-104. 

[3] Qu Wenying, Xie Jianguo, Liang Anxin & Shang Xuefeng. 
(2021). Effects of astragalus polysaccharides on nerve cell 
activity, cognitive function and Caspase-9 expression levels in 
rats with Alzheimer's disease. Stroke and Neurological 
Diseases (05), 543-549. 

[4] MENG Kaitao, ZHANG Jianguo, LIU Chong, MA Xiuhong & 
SONG Yuqiang. (2021). Correlation analysis between serum 
miR-128 and miR-223 expression levels and inflammatory 
response and cognitive function in patients with Alzheimer's 
disease of different severity. Stroke and Neurological Diseases 
(06), 667-671.