Random Forest Classification of Cognitive Impairment Using Digital

Tree Drawing Test (dTDT) Data

Sebastian Unger

, Zafer Bayram

, Laura Anderle

and Thomas Ostermann

Department of Psychology and Psychotherapy, Witten/Herdecke University,

Alfred-Herrhausen-Str. 50, 58448 Witten, Germany

Department of Informatics and Communication, Westphalian University of Applied Science,

Neidenburger Str. 43, 45897 Gelsenkirchen, Germany

Keywords: Digital Tree Drawing Test, Cognitive Impairment, Mental Disorders, Classification, Random Forest.

Abstract: Early detection and diagnosis of dementia is a major challenge for medical research and practice. Hence, in

the last decade, digital drawing tests became popular, showing sometimes even better performance than their

paper-and-pencil versions. Combined with machine learning algorithms, these tests are used to differentiate

between healthy people and people with mild cognitive impairment (MCI) or early Alzheimer's disease (eAD),

commonly using data from the Clock Drawing Test (CDT). In this investigation, a Random Forest

Classification (RF) algorithm is trained on digital Tree Drawing Test (dTDT) data, containing socio-medical

information and process data of 86 healthy people, 97 people with MCI, and 74 people with eAD. The results

indicate that the binary classification works well for homogeneous groups, as demonstrated by a sensitivity

of 0.85 and a specificity of 0.9 (AUC of 0.94). In contrast, the performance of both binary and multiclass

classification degrades for groups with heterogeneous characteristics, which is reflected in a sensitivity of

0.91 and 0.29 and a specificity of 0.44 and 0.36 (AUC of 0.74 and 0.65), respectively. Nevertheless, as the

early detection of cognitive impairment becomes increasingly important in healthcare, the results could be

useful for models that aim for automatic identification.

1 INTRODUCTION

Early detection and diagnosis of dementia, especially

in its early stages, is a major challenge in medical

research and practice (Yamasaki & Ikeda, 2024).

Traditional methods such as Shulman's Clock

Drawing Test (CDT) have proven useful for detecting

moderate to severe dementia but show limitations in

identifying mild cognitive impairment (MCI, Huang

et al., 2023).

In this context, digital drawing tests have become

more popular. By using a tablet and a pressure-

sensitive stylus, patients are asked to create drawings

on a tablet, which requires a complex interplay of

different cognitive abilities. Examples of such

drawing tests (see Figure 1 and Figure 2) include the

CDT (CDT, Yuan et al., 2021), the Spiral Drawing

Test (SDT, Fujiwara et al., 2023), and the digital Tree

Drawing Test (dTDT, Robens et al., 2019). The

benefits of those tests are that they create a less

stressful situation for the patient through creative

image design and freer presentation options, but also

use modern software for data collection, evaluation,

and statistical analyses of the complete drawing

process. This opens the potential to determine the

severity of dementia from a more patient-oriented

perspective and to enable an art-based but at the same

time reliable screening for patients with MCI and

early-stage Alzheimer's disease (eAD).

Figure 1: Examples of the Clock Drawing Test (left; Yoon

& Ahn, 2023) and the Spiral Drawing Test (right; Müller et

al., 2017).

Unger, S., Bayram, Z., Anderle, L. and Ostermann, T.

Random Forest Classiﬁcation of Cognitive Impairment Using Digital Tree Drawing Test (dTDT) Data.

DOI: 10.5220/0012859100003756

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Data Science, Technology and Applications (DATA 2024), pages 585-592

ISBN: 978-989-758-707-8; ISSN: 2184-285X

585

Thus, digital drawing tests not only have become

a promising tool in different areas of health services,

but also show comparable or sometimes even superior

performance than their paper-and-pencil versions,

according to recent systematic reviews and meta-

analyses of screening studies on MCI (Chan et al.,

2021; Ding et al., 2022). At the same time, the

digitization of these tests enables the combination of

process data with machine learning algorithms.

Currently, the CDT data is commonly used for

classification tasks (Binaco et al., 2020; Jimenez-

Mesa et al., 2022). The same applies when dealing

with SDT data (Akyol, 2017; Fahim et al., 2021).

However, with regard to the dTDT data, only a few

approaches exist, which could be due to the

complexity of data. While drawing tests such as the

CDT or the SDT mainly focus on graphomotoric

aspects (i.e., drawing movement), processing aspects

(i.e., time of completion and speed), and spatial

reasoning (deviation from a given form), the dTDT

also includes texture features (e.g., the use and change

of colors or stroke width).

Results, which were obtained by using a logistic

regression model (Robens et al, 2019), show that the

research with dTDT data is worth continuing: Firstly,

patients, suffering on cognitive impairments, have a

tendency to draw smaller and simpler images, which

were not positioned centrally on the drawing surface.

Secondly, a lack of variety in the selection of colors

and line widths was observed, which could indicate a

limited creative decision-making ability. Thirdly, the

movements of the pencil were less fluid and less

coordinated, sometimes even fleeting with a tendency

towards increased movements in the air. And finally,

a reduced speed when drawing, a delayed start to the

drawing process, and longer pauses when not drawing

were observed.

This investigation aims to extend the analyses of

Robens et al. (2019) by using a Random Forest

Classification (RF) model for predicting cognitive

impairment. RF was chosen because it already proved

to be adequate for handling dTDT data when applied

in binary classification models (Li et al., 2022). In

contrast to these two previous studies, not only the

binary classification is investigated here, but also a

first step towards a multiclass classification. Such

multiclass model would be beneficial for the

classification of impairments in clinical practice, as

the prediction would not depend on the model

selected according to the given circumstances, i.e.,

the experience of the medical professional and the

symptoms of a patient.

2 MATERIAL AND METHODS

2.1 Dataset

The dataset initially contains 66 numeric features of

257 people who were asked to draw a tree, similar to

Koch's tree test (Koch, 1952). In contrast to Koch's

tree test, these people had to draw the tree digitally

and were not bound by the requirement to draw a fruit

tree (see Figure 2 for an example).

The process data recorded during the drawing

make up the majority of the initial features. Other

features include socio-medical information, i.e., age,

gender, and the score of the Mini-Mental Status

Examination (MMSE) questionnaire. The feature

describing the people's cognitive health condition

assessed by medical professionals is used as outcome.

With this, the people can be divided into three groups:

a healthy control group (HC, 86 people), a group with

MCI (97 people), and a group with eAD (74 people).

The MCI and eAD group can also be viewed as a

combined group: the cognitive impaired group, which

is the opposite of the healthy control group (nonHC,

171 people).

Figure 2: Example of a digital tree drawing taken from

(Robens & Ostermann, 2020).

The socio-medical information of the three

groups is given in Table 1, revealing some significant

differences. Looking at the gender balance, male

participants are dominant in the HC group, while

female participants are the majority of the other two

groups, ranging from 53.1 % in the MCI group to

70.3 % in the eAD group. There are also imbalances

between the groups with respect to age and

educational years. Patients in the nonHC group were

significantly older than those in the HC group.

Moreover, participants in the HC group had more

educational years (14.0 years) than those in the MCI

(12.9 years) and eAD group (11.1 years).

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

586

To enhance the dataset and compensate for the

imbalances, further features, e.g., image colors or

texture characteristics, were extracted from the tree

images. These features were taken from the findings

of previous studies, investigating cognitive condition

(Ostermann et al., 2020; Robens et al., 2020). In

addition, features that are easy to calculate were

added, e.g., image size, ratio between image and

screen, or center of mass. At the end, there were a

total of 22 new features that, together with the others,

form the basis for possible predictors.

Table 1: Socio-medical information of the participants

subdivided by their cognitive health condition (MMSE:

Mini-Mental Status Examination; M: Mean; SD: Standard

deviation; *: significant differences between the groups).

HC MCI eAD

Number 86 97 74

Gender*

34 (39.5 %)

52 (60.5 %)

52 (53.6 %)

45 (46.4 %)

52 (70.3 %)

22 (29.7 %)

Female

Male

64.9 ± 10.4

68.1 ± 12.0

73.6 ± 11.1

M ± SD

Median

Education*

14.0 ± 3.0

12.9 ± 2.8

11.1 ± 3.1

M ± SD

Median

MMSE*

29.2 ± 0.9

26.3 ± 2.1

22.1 ± 3.0

M ± SD

Median

2.2 Process Flow

The process flow for developing the binary and

multiclass models consisted of four steps, whereby

the dataset with the initial features served as the basis

(Figure 3). First, the data set was prepared for training

the models. This was followed by the selection of the

relevant features. Once these were determined, the

models were optimized by tuning the

hyperparameters. Finally, the models were evaluated

with unseen test data.

Since feature selection (except pre-selection) and

model evaluation are intentionally randomized steps,

they were repeated 20 times. Only the split of the

dataset was controlled at the pre-processing step by

setting a seed, which shall lead to comparable results.

2.2.1 Data Pre-Processing

First, the dataset was checked for missing values. If

values were missing, the person and all according

data were removed from the dataset. After that, the

features were normalized so that the models could

process the dataset better. Normalization can be done

with various methods. This investigation uses a min-

max approach, which achieves comparatively good

results by rescaling the features into a new range of

values (Jayalakshmi & Santhakumaran, 2011). The

following formula expresses the used scaling:

𝑍

,



𝑋

,

 min



𝑋





max



𝑋





 min𝑋





)

With this scaling, each value X

i,j

of a feature i

becomes a value Z

i,j

between 0 and 1. The functions

min and max denote the minimum and maximum

values of feature i.

After normalization, the dataset was divided into

two parts using a fixed seed. The first part was used

to train the models and the second part to test the

models. The ratio of training data and test data was

80:20. All the subsequent steps were performed with

this dataset split to ensure that the same data is always

used for training and that the data for evaluation has

never been seen before.

Figure 3: Process flow of model development, starting with

the dataset itself, through pre-selection to evaluation.

2.2.2 Feature Selection

In the first step of feature selection, the 66 initial

features were pre-selected to a total of 19 using a list

created by art therapists specialized in dementia

Random Forest Classiﬁcation of Cognitive Impairment Using Digital Tree Drawing Test (dTDT) Data

587

(Robens et al., 2019). The 22 newly added features

remained unaffected by this reduction.

In the second step, a recursive feature elimination

approach with a 10-fold cross-validation (RFECV)

were used. The RF algorithm wrapped by the RFECV

were used to determine the features within the

training data. The approach identifies relevant

features by using all possible feature combinations

- 1), starting with all features and successively

reducing the number of features. The relevance was

then indicated by ascending numbers, whereby

relevant features that have been selected were

marked.

In a third step, the features from the second step

were further viewed in accordance with their

contribution to the model’s accuracy, which was

indicated by the number of selected features in

addition to their ranking. Therefore, a feature was

eliminated if it only contributes with many other

features. This was to ensure that the most relevant

features (frequent occurrence plus high rank) were

identified as predictors for the classification task.

2.2.3 Hyperparameter Tuning

After the most relevant features were selected, the

hyperparameters of the RF models for the binary and

multiclass classification had to be tuned to further

improve the models’ accuracy. This was done using a

grid search approach, again, with 10 folds for cross-

validation (GSCV).

Seven hyperparameters (n_estimators,

max_depth, min_samples_split, min_samples_leaf,

max_features, bootstrap, and class_weight) were

tuned, starting with a wide range of values. The range

then was optimized step by step until no further

improvement in accuracy could be observed.

2.2.4 Evaluation

For the evaluation, the RF models were also trained

with the 10-fold GSCV. After training, the RF models

received the test data the first time to perform their

prediction.

Accuracy, precision, sensitivity, specificity, and

F1-score were used to assess the performance of the

models. Moreover, the diagnostic power of the

selected features was analyzed using the areas under

curve (AUC). The interpretation is as follows (Polo

& Miot, 2020):

 worthless: 0.6 - 0.7;

 poor: 0.7 - 0.8;

 good: 0.8 - 0.9;

 excellent: > 0.9.

3 RESULTS

3.1 Data Pre-Processing

When checking the dataset for missing values, only

one value and therefore one person was removed

from the dataset, leaving 256 people for the model

development. The subsequent split of the dataset into

training and test data resulted in 204 and 52 people,

respectively.

3.2 Feature Selection

The visual output (Figure 4) of the RFECV indicates

that the optimal number of features for all three

models is probably in the lower decimal range. From

a value of around 15, all models appear to stagnate,

which could indicate overfitting. Therefore, between

10 and 15 individual features seems to be optimal for

each model.

When looking at the quantitative output of the

RFECV, similarities can be observed. 15.45 ± 8.48

features are used to classify the HC and nonHC

groups, 23.35 ± 7.71 for HC and eAD groups, and

14 ± 4.95 for HC, MCI, and eAD groups. Except for

the classification of HC and eAD, the optimal number

of features falls within the previously assumed range

due to the standard deviation.

Table 2: Features remaining for model evaluation (B1:

binary classification of HC and nonHC; B2: binary

classification of HC and eAD; M: multiclass classification

of HC, MCI, and eAD).

Feature B1 B2 M

Color Chan

es X X

Color Count X

Contrast X X

Duration (ms) X X

Image Width X X X

Not Paintin

(

)

X X X

e Relation X X

Paintin

(

)

Pen Up Count X X X

Pen Up Pen Down Relation X X X

Pen Up Stroke Length X X

Pressure Velocit

Relation X X X

Stroke Chan

es X X

Strokes Per Minute X X X

Velocity Mean X X X

Volatile Motion Mean X X X

All features considered relevant are listed in Table 2.

There are 12 features for the two binary classification

models and 14 for the multiclass classification model.

Specific colours (e.g., red, green, or yellow) and the

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

588

center of mass of the image (i.e., x and y coordinate

of the pixel) were excluded in the feature selection

process mainly because they were only partially used

to train the RF models and had less importance when

they appeared. Moreover, socio-medical information,

i.e., age, educational years, gender, and other mental

health related data such as MMSE or data from CDT,

were also excluded.

Figure 4: Results of the RFECV. For each model, the

diagrams show the correlation between the number of

features and their corresponding accuracy.

3.3 Hyperparameter Tuning

Tuning the hyperparameter resulted in a significant

improvement in accuracy for each model. It was best

with the multiclass model. Its accuracy during feature

elimination was 0.49 ± 0.04, with a peak of about

0.59 after tuning. The two binary classification

models showed a similarly good improvement. The

accuracy increased from 0.68 ± 0.02 to about 0.74 in

the classification of HC and nonHC and from

0.75 ± 0.04 to about 0.82 in the classification of HC

and eAD.

3.4 Evaluation

The binary model to classify HC and nonHC showed

the second best results. It had a mean accuracy of

74 % and was quite successful at detecting the non-

healthy people (sensitivity) but lacked in detecting the

healthy ones (specificity) as given in Table 3. Among

the most important features were “Velocity Mean”,

“Pen Up Count”, and “Strokes Per Minute”.

The distinction between HC and eAD group was

most successful. The model’s mean accuracy was

88 %. In detail, the model detected healthy people

similarly well as people with eAD, shown by a mean

specificity of 90 % and a mean sensitivity of 87 %,

respectively. Also here, “Velocity Mean” was one of

the most important features. In addition, “Color

Changes” and “Pressure Velocity Relation” had a

strong impact on the model.

Table 3: Metrics of the GSCV represented as mean and

standard deviation (B1: binary classification of HC and

nonHC; B2: binary classification of HC and eAD; M:

multiclass classification of HC, MCI, and eAD).

B1 B2 M

Accuracy 0.74 ± 0.02 0.88 ± 0.01 0.22 ± 0.01

Precision 0.74 ± 0.03 0.87 ± 0.01 0.4 ± 0.02

Sensitivit

0.91 ± 0.02 0.85 ± 0 0.29 ± 0.32

ecificit

0.44 ± 0.03 0.9 ± 0.02 0.36 ± 0.01

AUC 0.74 ± 0.01 0.94 ± 0 0.65 ± 0.01

F1-Score 0.68 ± 0.02 0.87 ± 0.01 0.29 ± 0.01

As presented in Table 3, the multiclass model was

the worst of the three, although “Velocity Mean” had

the most impact as in the two binary classifications.

In general, all features had a relatively equal impact

on the model, which was found by looking at their

importances. More importantly, the model showed

great difficulties in classifying the MCI group,

resulting in a mean accuracy of 22 %, which was even

below chance (33.3 %). This is similar in terms of

specificity and sensitivity.

Random Forest Classiﬁcation of Cognitive Impairment Using Digital Tree Drawing Test (dTDT) Data

589

Figure 5: ROC curve for binary and multiclass

classification. For each model, the diagrams show the

correlation between the sensitivity (true positive rate) and

the corresponding 1 – specificity (false positive rate).

Receiver operating characteristic (ROC) curves in

Figure 5 display the models’ performance for the

binary and multiclass classification. With mean

AUCs of 0.74 and 0.94 for the binary models, the

results can be considered as in need of improvement

and almost perfect, respectively. The mean AUC of

0.65 for the multiclass model is unfortunately not

sufficient.

4 DISCUSSION

This investigation describes the use of three RF

classification models for a dataset of healthy and

cognitive impaired people that completed the dTDT.

The features of the dataset were first reduced in

accordance with the similar study of Robens et al.

(2019). The remaining 19 features (22 including age,

gender, and educational years) were then expanded

by features that could easily be calculated and

features based on other findings of cognitive

condition (Ostermann et al., 2020; Robens et al.,

2020).

The evaluation of the trained models reveals that

the distinction between the HC group and the group

with eAD works quite well. The result is comparable

to previous studies on dTDT data (Li et al., 2022;

Robens et al., 2019), indicating a functional and valid

model. In contrast to that, no similar result could be

achieved with the distinction between the HC group

and the nonHC group. The model’s performance is

significantly worse as in the mentioned two studies,

but comparable with the result of a study on digitized

CDT data (Jimenez-Mesa et al., 2022).

Although the available features and selection

process were almost identical to the study of Robens

et al. (2019), the limiting factor here might be that the

dataset is too heterogeneous and too imbalanced. On

the one hand, mean MMSE scores between HC and

MCI and between MCI and eAD overlap (Table 1).

On the other hand, all three groups had similarly long

educational years. Since MMSE is a marker for

cognitive condition (Dellasega & Morris, 1993) and

education has a protective effect on developing

cognitive impairment such as AD (Sando et al.,

2008), the process data could be disturbed by these

circumstances, making a clear distinction not

possible. According to Wenner et al. (2020),

manipulating the training data and adjusting the

classifier could improve the classification with an

imbalanced dataset, which might be considered in

future studies.

For the low performance of the multiclass model,

which was below chance, the same limitations and

solutions mentioned for the binary classification

could be applied here. Another improvement might

be utilizing a model that specifically is designed for

the classification of trees based on their size

(Setiawan et al., 2020). Nevertheless, there is still a

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

590

need for further investigation, because even if it was

done with dCTD data, a prediction with a multiclass

classification can be better than by chance (Binaco et

al., 2020).

5 CONCLUSIONS

Early detection of cognitive impairment is an

increasingly important field in healthcare. Therefore,

the idea of combining machine learning algorithms

with digital drawing tasks to enable automatic

identification of cognitive impairments has been

explored for some time. With the here presented

results, which vary strongly depending on the

classification task, new insights could be provided for

handling dTDT data. Whereas the binary

classification of homogeneous and sufficiently

distinct groups works well, both binary and multiclass

classification seem to have their difficulties if the

characteristics that form a group are not distinct

enough.

ACKNOWLEDGEMENTS

This work was financially supported by a grant of the

Software AG Foundation, Darmstadt, Germany.

REFERENCES

Akyol, K. (2017). A study on the diagnosis of Parkinson’s

disease using digitized wacom graphics tablet dataset.

Int J Inf Technol Comput Sci, 9, 45-51.

Binaco, R., Calzaretto, N., Epifano, J., McGuire, S., Umer,

M., Emrani, S., Wasserman, V., Libon, D. J., & Polikar,

R. (2020). Machine learning analysis of digital clock

drawing test performance for differential classification

of mild cognitive impairment subtypes versus

Alzheimer’s disease. Journal of the International

Neuropsychological Society, 26(7), 690-700.

Chan, J. Y., Bat, B. K., Wong, A., Chan, T. K., Huo, Z.,

Yip, B. H., Kowk, T.C.Y., & Tsoi, K. K. (2021).

Evaluation of digital drawing tests and paper-and-

pencil drawing tests for the screening of mild cognitive

impairment and dementia: a systematic review and

meta-analysis of diagnostic studies. Neuropsychology

Review, 1-11.

Dellasega, C., & Morris, D. (1993). The MMSE to assess

the cognitive state of elders. Journal of Neuroscience

Nursing, 25(3), 147-152.

Ding, Z., Lee, T. L., & Chan, A. S. (2022). Digital cognitive

biomarker for mild cognitive impairments and

dementia: a systematic review. Journal of clinical

medicine, 11(14), 4191.

Fahim, M. I., Islam, S., Noor, S. T., Hossain, M. J., & Setu,

M. S. (2021). Machine learning model to analyze

telemonitoring dyphosia factors of Parkinson’s disease.

International Journal of Advanced Computer Science

and Applications, 12(8).

Fujiwara, K., Matsuhashi, K., & Mitobe, K. (2023).

Detection of Mild Cognitive Impairment Using a Spiral

Drawing Test. Journal of Advanced Computational

Intelligence and Intelligent Informatics, 27(5), 907-

914.

Huang, Y., Pan, F. F., Huang, L., & Guo, Q. (2023). The

value of clock drawing process Assessment in

Screening for mild cognitive impairment and

Alzheimer’s dementia. Assessment, 30(2), 364-374.

Jayalakshmi, T., & Santhakumaran, A. (2011). Statistical

normalization and back propagation for classification.

International Journal of Computer Theory and

Engineering, 3(1), 1793-8201.

Jimenez-Mesa, C., Arco, J. E., Valentí-Soler, M., Frades-

Payo, B., Zea-Sevilla, M. A., Ortiz, A., Ávila-

Villanueva, M., Castillo-Barnes, D., Ramírez, J., del

Ser-Quijano, T., Carnero-Pardo, C., & Górriz, J. M.

(2022). Automatic classification system for diagnosis

of cognitive impairment based on the clock-drawing

test. In International Work-Conference on the Interplay

Between Natural and Artificial Computation (pp. 34-

42). Cham: Springer International Publishing.

Li, J., Yang, J., Yang, J., Yang, H., Lan, M., & Gao, L.

(2022, November). Characterizing cognitive

impairment through drawing features extracted from

the Tree Drawing Test. In 2022 7th International

Conference on Intelligent Informatics and Biomedical

Science (ICIIBMS) (pp. 341-347). IEEE.

Koch, C. (1952). The Tree Test: The Tree-drawing Test as

an Aid in Psychodiagnosis. Berne: Hans Huber

Publishing

Müller, S., Preische, O., Heymann, P., Elbing, U., & Laske,

C. (2017). Increased diagnostic accuracy of digital vs.

conventional clock drawing test for discrimination of

patients in the early course of Alzheimer’s disease from

cognitively healthy individuals. Frontiers in aging

neuroscience

, 9, 101.

Ostermann, T., Robens, S., Heymann, P., Unger, S., Müller,

S., Laske, C., & Elbing, U. (2020). Analysis of the Use

of Colour for Early Detection of Dementia. In

Proceedings of the 13th International Joint Conference

on Biomedical Engineering Systems and Technologies

(BIOSTEC 2020) – HEALTHINF (pp. 316-320).

Polo, T. C. F., & Miot, H. A. (2020). Use of ROC curves in

clinical and experimental studies. Jornal vascular

brasileiro, 19, e20200186.

Robens, S., Heymann, P., Gienger, R., Hett, A., Müller, S.,

Laske, C., Loy, R., Ostermann, T., Elbing, U. (2019).

The digital tree drawing test for screening of early

dementia: an explorative study comparing healthy

controls, patients with mild cognitive impairment, and

patients with early dementia of the Alzheimer type.

Journal of Alzheimer's Disease, 68(4), 1561-1574.

Random Forest Classiﬁcation of Cognitive Impairment Using Digital Tree Drawing Test (dTDT) Data

591

Robens, S., & Ostermann, T. (2020). Der digitale

Baumzeichentest–Ein kunsttherapeutischer Ansatz für

das Demenz-Screening. Zeitschrift für

Komplementärmedizin, 12(05), 24-28.

Robens, S., Ostermann, T., Heymann, P., Müller, S., Laske,

C., & Elbing, U. (2020). Comparison of texture features

and color characteristics of digital drawings in

cognitive healthy subjects and patients with amnestic

mild cognitive impairment or early alzheimer’s

dementia. In Biomedical Engineering Systems and

Technologies: 12th International Joint Conference (pp.

412-428). Springer International Publishing.

Sando, S. B., Melquist, S., Cannon, A., Hutton, M.,

Sletvold, O., Saltvedt, I., White, L. R., Lydersen, S., &

Aasly, J. (2008). Risk‐reducing effect of education in

Alzheimer's disease. International Journal of Geriatric

Psychiatry: A journal of the psychiatry of late life and

allied sciences, 23(11), 1156-1162.

Setiawan, I., Yusnitasari, T., Nurhady, H., & Hizviani, N.

V. (2020). Implementation of convolutional neural

network method for classification of Baum Test. In

2020 fifth international conference on informatics and

computing (ICIC) (pp. 1-6). IEEE.

Wenner, M., Hibert, C., Meier, L., & Walter, F. (2020).

Near real-time automated classification of seismic

signals of slope failures with continuous random

forests. Natural Hazards and Earth System Sciences

Discussions, 2020, 1-23.

Yamasaki, T., & Ikeda, T. (2024). Advances in Research on

Brain Health and Dementia: Prevention and Early

Detection of Cognitive Decline and Dementia. Brain

Sciences, 14(4), 353.

Yoon, H., & Ahn, M. (2023). Quantification of Movement

Error from Spiral Drawing Test. Sensors, 23(6), 3043.

Yuan, J., Libon, D. J., Karjadi, C., Ang, A. F., Devine, S.,

Auerbach, S. H., Au, R., & Lin, H. (2021). Association

between the digital clock drawing test and

neuropsychological test performance: large

community-based prospective cohort (Framingham

heart study). Journal of Medical Internet Research,

23(6), e27407.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

592