Machine Learning Approaches for Diabetes Classiﬁcation:

Perspectives to Artiﬁcial Intelligence Methods Updating

Giuseppe Mainenti

, Lelio Campanile

, Fiammetta Marulli

2,∗

, Carlo Ricciardi

and Antonio S. Valente

DIEM, University of Salerno, Salerno, Italy

Dep. di Matematica e Fisica, Universit

a Della Campania ”L. Vanvitelli”, Caserta, Italy

Department of Advanced Biomedical Sciences, University of Naples ”Federico II”, Naples, Italy

Public Funding Accountable, San Nicola La Strada, Caserta, Italy

antonio.valente@ak12srl.it

Keywords:

Diabetes Classiﬁcation, Diabetes Management, Machine Learning, Artiﬁcial Intelligence, Big Data Analytics.

Abstract:

In recent years the application of Machine Learning (ML) and Artiﬁcial Intelligence (AI) techniques in health-

care helped clinicians to improve the management of chronic patients. Diabetes is among the most common

chronic illness in the world for which often is still challenging do an early detection and a correct classiﬁ-

cation of type of diabetes to an individual. In fact it often depends on the circumstances present at the time

of diagnosis, and many diabetic individuals do not easily ﬁt into a single class. The aim is this paper is the

application of ML techniques in order to classify the occurrence of different mellitus diabetes on the base of

clinical data obtained from diabetic patients during the daily hospitals activities.

1 INTRODUCTION

Recent estimation accounts about 460 million people

worldwide affected by diabetes, and yet 1 in 2 per-

sons remain untreated or undiscovered, causing blind-

ness, ﬁngers amputations, kidney failures and almost

doubling the risk of heart attack and all-cause mor-

tality, leading to hospitalization, long-term complica-

tions, and higher costs also for healthcare infrastruc-

tures. According to International Diabetes Federation

previsions, for the next 20 years, the number of peo-

ple affected by diabetes will reach about 700 million

(IDF, 2017).

In (Patterson et al., 2019) methods, results and

limitations of the 2019 International Diabetes Feder-

ation (IDF) Diabetes Atlas 9th edition are described,

thus providing an estimation of worldwide numbers

of cases of type 1 diabetes in children and adoles-

cents. The performed research provided as insights

that incidence rates were available for 45% of coun-

tries, ranging from 6% in the sub-Saharan Africa re-

gion to 77% in the European region), thus concluding

that the worldwide estimation for the number of chil-

dren and adolescents with type 1 diabetes continue to

∗

Corresponding Author.

increase.

In this paper we experimentally analyze the adop-

tion of AI in medicine with the aim to investigate how

it could improve accuracy of diagnosis, making life

easier to patients and clinicians. The application of

AI will open the development of new care paradigms,

aiding peoples in the administration of drugs, as well

as in the implementation of personal medical assis-

tants in every smartphone. The paper is organized as

follows: Section 2 describes related works and the

topic related literature; Section 3 explains the pro-

posed research, describing the methods, the classiﬁ-

cation algorithms performed and the motivating case

study. Section 4 describes the experimental design

and setting as the data set and the evaluation metrics.

Section 5 discusses the results from the performed ex-

periments and Section 6 explains conclusions and fu-

ture improvements for the presented research.

2 RELATED WORKS

Applications of ML, AI and cognitive computing

(CC) offer effective promises in healthcare domain.

The early detection of most of health diseases is cru-

Mainenti, G., Campanile, L., Marulli, F., Ricciardi, C. and Valente, A.

Machine Learning Approaches for Diabetes Classiﬁcation: Perspectives to Artiﬁcial Intelligence Methods Updating.

DOI: 10.5220/0009839405330540

In Proceedings of the 5th International Conference on Internet of Things, Big Data and Security (IoTBDS 2020), pages 533-540

ISBN: 978-989-758-426-8

533

cial to prevent tragic consequences and automatic sys-

tems and algorithms provide an effective support for

their diagnosis. Diabets represents one of the most

spread and ever increasing disease affecting the world

population, as also for the carotid disease.

The study provided in (Verde and De Pietro, 2018)

investigates several machine learning techniques able

of detecting the presence of a carotid disease; such

study compares performances of different machine

learning methods by analysing the Heart Rate Vari-

ability (HRV) parameters of electrocardiographic sig-

nals selected from an on line available database on

the Physionet website. The results obtained by the

comparative study are provided in terms of accuracy,

precision, recall and F-measure metrics.

Nowadays, the adoption of mobile technologies in

the healthcare sector is also increasing signiﬁcantly,

since they provide promising solutions for people

who desire the detection, monitoring, and treatment

of their health conditions anywhere and at any time,

beyond offering ways for communicating multimedia

content (e.g. clinical audio-visual notes and medi-

cal records). This aspect is investigated in (Verde

et al., 2018) where several machine learning tech-

niques are compared for supporting investigations

on voice pathology and disorders detection, refer-

ring speciﬁcally to the dysphonia, an alteration of the

voice quality affecting about one person in three at

least once in his/her lifetime. The results of the per-

formed study are provided in terms of accuracy, sen-

sitivity, speciﬁcity, and receiver operating character-

istic area. Alternative algorithms, as support vector

machine or decision tree, are ﬁnally suggested, de-

pending on the features evaluated and feature selec-

tion methods.

As for diabetes, ML and AI applications and

researches are quickly growing even more, as it

is evidenced by the estimation in the number

of related articles indexed in the Google Scholar

database(Contreras and Veh

ı, 2018).

In (Dankwa-Mullan et al., 2019) is performed an

accurate literature analysis, addressing currently AI

advances relevant to diabetes ecosystem, made of pa-

tients affected by diabetes, families, caregivers and

clinicians. Such literature recognition was performed

by consulting the online database PubMed (Fiorini

et al., 2017). Key terms adopted for research were

“diabetes” and “artiﬁcial intelligence”, by which they

met 450 publicly available articles and sources of in-

formation from 2009 to 2019.

Furthermore, global market insights related to the

2019 are analyzed in (Zhou and Myrzashova, 2020),

and research results reveal that an explosive growth of

40% is expected from 2017 to 2024 in the use of AI

in the health care ﬁeld. The use of AI in medicine to-

day can signiﬁcantly improve the accuracy of diagno-

sis, make life easier for patients with various diseases,

and with the development of technology it will make

the emergence of highly effective personal medicines

real, as well as a personal medical assistant in every

smartphone. Nowadays, another clinical challenge

is the realization of personalized treatments, which

falls into the more general paradigm of the precision

medicine.

In (Coronato and Naeem, 2019) a Reinforcement

Learning paradigm is adopted to build an intelligent

system, able to self-adapt to users’ skills aiming at

assisting them in the healthcare treatment. A method

to classify patients affected by diabetes using a set of

characteristic selected in according to World Health

Organization criteria is described in (Mercaldo et al.,

2017), where the state of the art ML algorithms were

applied for valuating real-world data. Experiments

obtained a 0.770 value for the precision metric and

a recall equal to 0.775 using the HoeffdingTree al-

gorithm. (Brunese et al., 2019) proposes a neural

network-based method aimed to discriminate between

different lung cancer types, to assist medics and radi-

ologists in the diagnosis formulation. In this work,

they adopt a set of 30 radiomic feature directly ob-

tained from magnetic resonance; the neural network

model is tuned during the variation of the momentum

and the loss functions and was evaluated on a data set

consisting of 2000 MRI labelled through medical re-

ports.

Researches in (Osman and Aljahdali, 2017) pro-

posed an integration approach between the SVM tech-

nique and K-means clustering algorithms to diag-

nose diabetes disease. The focus of the method was

adjusted so that only the most important features

received attention. They performed the T-Test or-

der to quantify the improvements achieved by their

approach before and after combination process be-

tween K-means and SVM algorithms. Authors in

(Nanda et al., 2011) develop a model for the pre-

diction of gestational diabetes mellitus (GDM) from

maternal characteristics and biochemical markers at

11 to 13 weeks’ gestation. They demonstrate that in

the screening study, maternal age, body mass index,

racial origin, previous history of GDM and macro-

somic neonate were signiﬁcant independent predic-

tors of future. The detection rate was 61.6% at a

false-positive rate of 20% and the detection increased

to 74.1% by the addition of adiponectin and sex

hormone-binding globulin. A data informed frame-

work for identifying subjects with and without Type

2 Diabetes Mellitus (T2DM) from Electronic Health

Records via feature engineering and ML is proposed

AI4EIoTs 2020 - Special Session on Artiﬁcial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives

534

in (Zheng et al., 2017). K-means clustering analysis is

adopted in (Bennetts et al., 2013) in order to identify

typical regional peak plantar pressure distributions in

a group of 819 diabetic feet. The number of analyzed

clusters ranged from 2 to 10 to examine the effect on

the differentiation and classiﬁcation of regional peak

plantar pressure distributions. Such analysis aimed to

provide an understanding of the variability of the re-

gional peak plantar pressure distributions seen within

the diabetic population and serves as a guide for the

preemptive assessment and prevention of diabetic foot

ulcers. Finally, in (Alssema et al., 2011) authors em-

ployed the Finnish diabetes risk questionnaire to iden-

tify subjects exposed at risk for drug-treated Type 2

diabetes. in this research, additional predictors were

added and the risk questionnaire was updated by us-

ing clinically diagnosed and screen-detected Type 2

diabetes instead of drug-treated diabetes. The con-

clusion of the study was that the predictive value of

the original Finnish risk questionnaire could be im-

proved by adding information on sex, smoking and

family history of diabetes.

In (Brunese et al., 2020) a research aiming to rec-

ognize the different brain cancer grades by analysing

brain magnetic resonance images, since the brain can-

cer is one of the most aggressive tumour. The pro-

posed method aims to identify the components of an

ensemble learner. The ensemble learner is focused

on the discrimination between different brain cancer

grades using non invasive radiomic features, belong-

ing to ﬁve different groups: First Order, Shape, Gray

Level Co-occurrence Matrix, Gray Level Run Length

Matrix and Gray Level Size Zone Matrix. As re-

search results, authors evaluated the features effec-

tiveness through hypothesis testing and through deci-

sion boundaries, performance analysis and calibration

plots thus we select the best candidate classiﬁers for

the ensemble learner.

3 THE PROPOSED RESEARCH

Computerization in healthcare is on the rise thus lead-

ing to large patient databases, with speciﬁc proper-

ties. ML techniques are able to examine and to ex-

tract knowledge from large databases in an automatic

way (Meyfroidt et al., 2009). Several examples can

be found in literature as regards the classiﬁcation of

diagnosis or prognosis of different pathologies in a

wide range of medical ﬁelds (Ricciardi et al., 2020a),

(Ricciardi et al., 2020b), (Romeo et al., 2020).

In this work a number of predictive models was

implemented including Decision Tree, Random For-

est, k-Nearest Neighbors (KNN), Logistic Regres-

sion and Multilayer Perceptron to classify the type

of diabetes of the patients. Similar approach have

been investigated in literature in order to classify car-

diotocogram data (Sahin and Subasi, 2015). To con-

duct our analysis we exploited data of patients con-

tained in a clinical database that has been made avail-

able within the TablHealth Project.

After analyzing the data and selecting the most

suitable subset, a Data Cleaning strategy has been ap-

plied to manage the imbalance between classes and

missing data. Based on the dataset obtained, a web

platform was then implemented using Flask, a web

micro-framework written in Python, to classify ﬁve

different types of diabetes: type 1 diabetes, type 2 di-

abetes, gestational diabetes, diabetes Maturity-Onset

Diabetes of the Young (MODY) and diabetes Latent

Autoimmune Diabetes of Adult (LDA). The cate-

gorical variables considered were: age, diet ([kcal]),

fasting blood sugar, diastolic blood pressure, sys-

tolic blood pressure, Body Mass Index (BMI), gly-

cated hemoglobin (HbA1c), reduced glucose toler-

ance, syndrome metabolic disease, macrosomy, mi-

croalbuminuria, ischemic heart disease, high blood

pressure and cerebral vasculopathy.

3.1 Automatic Classiﬁcation of Diabetes

Types

The idea of the present work is that to exploit ML

techniques to investigate and build classiﬁcation mod-

els that can predict with high accuracy the type of dia-

betes. The search for a better prediction and therefore

a faster diagnosis by the specialized medical staff can

be vital for the health of the diabetic patient, forced to

face numerous complications that can become lethal

if not managed in a timely manner.

3.2 The Methodology

In Python, the following algorithms have been im-

plemented through the use of the Scikit-Learn frame-

work:

• Decision Tree: a structure made up of leaves and

nodes that represent the attributes and the pre-

dicted classes;

• Random Forest: it is a results of applying ran-

domization and bagging on the decision tree to

improve its accuracy;

• KNN: it is an instance-based model;

• Logistic Regression: it is a statistical-based

model, usually employed for binary classiﬁcation;

Machine Learning Approaches for Diabetes Classiﬁcation: Perspectives to Artiﬁcial Intelligence Methods Updating

535

Figure 1: Classiﬁcation process.

• Multilayer Perceptron: this is an example of neu-

ral network, a complex structure simulating the

one of the human brain.

3.3 The Case Study

The idea behind it was to provide the specialized med-

ical technical staff with an interface that would allow

all features that were considered to be in input when

creating the dataset and then in the training of the

models, choose The model to be used and output is

shown the resulting prediction made on the input data

with the chosen model.

4 EXPERIMENT DESIGN

A ML classiﬁcation algorithm to make the best needs

a good dataset. The creation of a well-structured

dataset suitable for the algorithm is far from trivial.

In this paper work the data was made available under

courtesy from the company where I did the internship.

They were collected in an Excel ﬁle by medical staff

from patients who made outpatient visits. We can di-

vide the process of implementing and developing the

web application into three phases:

1. Creating the dataset;

2. Implementing modeling algorithms;

3. Web application development with Flask.

4.1 Data Set

The manipulation of the dataset used for the algo-

rithms, has been done in Jupyter Notebook environ-

ment by means the Pandas library.

Table 1: Features.

Parameter [Unit] Type Range

Age [Year] Integer [11,92]

Diet [Kcal] Float [1200, 2000]

Fasting blood sugar [mg/dL] Float [52, 434]

Systolic blood pressure [mmHg] Float [90, 180]

Diastolic blood pressure [mmHg] Float [60, 95]

BMI [kg/m

] Float [18.25, 63.19]

Glycated hemoglobin [HbA1c%] Float [4.71, 14.40]

Reduced glucose tolerance Integer {0,1}

Syndrome metabolic disease Integer {0,1}

Macrosomy Integer {0,1}

Microalbuminuria [mg/L] Float [10, 300]

Ischemic heart disease Integer {0,1}

High blood pressure Integer {0,1}

Cerebral vasculopathy Integer {0,1}

Step 1. A data cleaning and analysis has been ex-

ecuted on the information contained into the Excel

sheet in which were registered all patient data and

output classes. To this aim data have been imported

into a dataframe and analyzed thanks to command

dataframe.info() of Pandas. The result of the analy-

sis showed several errors, such as missing, noisy and

inconsistent data. Therefore, we proceeded to clean

and normalize data by exploiting the Python’s Pandas

library in order to get a good dataset. At the end of

this ﬁrst step we obtained a dataset of 3378 elemnts

related to patients affected by diabetes.

Step 2. First of all we identiﬁed all the relevant fea-

tures such as BMI, fasting blood glucose, etc. We

found 14 valid features, reported in Table 1 with the

type and the range for the sake of clarity.

Then, ﬁve output classes distinguishing the type

of diabetes have been identiﬁed:

• Type 1 diabetes;

• Type 2 diabetes;

• Gestational diabetes;

• MODY diabetes;

• LADA diabetes;

Step 3. Finally a graphical analysis has been done to

emphasize the characteristics of dataset with regard-

ing the most signiﬁcative clinical features. In Figure

2 is depicted the boxplot of the BMI with respect the

Hb1c.

In 3 we reported the scatter-plot of fasting glucose

with respect BMI using a gradient scale of colors

AI4EIoTs 2020 - Special Session on Artiﬁcial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives

536

Figure 2: Boxplot for each outcome. On the x axis Bmi and

on the y axis Hb1c.

Figure 3: Frequency and fasting glucose features.

varying according to numbers of samples.

Note that the resulting dataset is characterized by a

strong imbalance, which fully reﬂects the state of the

art in literature. For proper classiﬁcation, the imbal-

ance of the dataset is a major obstacle. In these cases,

the classiﬁer used could pay special attention to the

majority class at the expense of the minority class.

Packages such as smoote are created to solve these

cases using under/over sampling techniques, however

they will not be used for this work. It is true that

the imbalance can affect the performance of the al-

gorithm, in terms of accuracy, precision and recall but

not timing performance, however if the imbalance af-

fects all the classiﬁers in the same way then the ab-

solute performance of the classiﬁer is affected but not

the relative one; Based on this, getting a good result

in these conditions comes from the robustness of the

classiﬁer compared to the imbalance.

4.2 Evaluation Metrics

The following evaluation metrics, well-known in liter-

ature for assessing the performance of ML algorithms

(Hossin and Sulaiman, 2015), were used:

• Accuracy: the number of correct predictions over

the total;

• Precision: the ratio between true positives and the

overall positives;

• Recall: the ratio between true positives and the

sum of true positive and false negatives;

• F1-score: an harmonic mean of precision and re-

call.

Each model has been evaluated with the metrics dis-

cussed above, logically related to the multiclass case,

in this case scikit-learn to evaluate non-binary clas-

siﬁcation models uses a normalized or macro-media.

The latter is simply calculated as the average scores

of different systems and weighs all classes equally

to evaluate a classiﬁer’s overall performance against

the most frequent class labels. The weighted macro-

average is calculated by weighing the score of each

label of the class by the number of true instances at

the time of the average calculation. It is useful if we

have a problem of poor class balance.

4.3 Implementing Classiﬁcation

Algorithms

Even at this stage the use of Jupyter Notebook was de-

cisive, because from time to time the code was modi-

ﬁed and executed in real time. In this section, we cre-

ate the models with the classiﬁcation algorithms listed

in Paragraph 3.2. The ﬁrst step taken at this stage is

to divide it into two parts, loading the dataset, and di-

viding the dataset into two parts, training datasets and

test datasets.

1. Loading the dataset with the pandas read csv()

method;

2. Assign features to a dataframe named x and

outputs to a y-named dataframe using the loc()

method of the Pandas library;

3. Separate the dataset into training datasets (80% of

the data) and test datasets (20% of the data), this

we can do with the train test split() method.

Scikit-Learn provides a class and various methods for

each template you want to use. The methods used to

perform the training and perform the prediction are

ﬁt() and predict(), respectively.

Machine Learning Approaches for Diabetes Classiﬁcation: Perspectives to Artiﬁcial Intelligence Methods Updating

537

4.3.1 Decision Tree

After importing the DecisionTreeClassiﬁer class, an

instance of the class itself was created and the ﬁt() and

predict() methods were executed. For this algorithm,

two models were tested, one with all the parameters

set by default and the other with the addition of the

parameter class weight = ”balanced”. This parame-

ter has been tested to cope with the imbalance of the

dataset, and as can be seen from the results obtained

(section 5) it is much better performing than the ﬁrst

model tested.

4.3.2 Random Forest

A random forest model can be considered as a set

of decision trees. The idea of putting more decision

trees together is to combine multiple simpler models

to build a more robust one that offers better classi-

ﬁcation performance and is less susceptible to over-

ﬁtting issues. The advantage of using random for-

est as a classiﬁer is that it doesn’t need to modify as

many hyper-parameters to achieve great performance.

After importing the RandomForestClassiﬁer class, an

instance of the class itself was created and the ﬁt()

and predict() methods were executed. As can be seen

from the result, the model is able to classify all out-

comes very well excluding the diabetes category lada

presumably because it is the category with the least

samples.

4.3.3 K-Nearest-Neighbors

For this model was created an object called ”knn”

that knows how to do KNN classiﬁcation once the

data is provided. All parameters of the KNeigh-

borsClassiﬁer class are set to the default values ex-

cluding the ”n neighbours” parameter, which is the

optimization/hyper-parameter (k) parameter. The

choice of the k value is of paramount importance

if you want to have the maximum accuracy of the

model, for this reason a method has been written that

does the train and test on the model with k ranging

from 0 to 25 and saves the result of accuracy, both

test and train, in a list and then plots the result with

the help of the matplotlib library. The choice of k = 1

is precisely the output of the function, where the ac-

curacy of the model was higher than the other values

of k. By evaluating the model with the selection of

all the parameters set by default and then with the op-

timization of the hyper-parameter k you can see how

all the evaluation metrics have improved.

4.3.4 Logistic Regression

Again, an instance of the LogisticRegression class

was created and the ﬁt() and predict() methods were

invoked on it. All parameters of the Logistic Regres-

sion class are set to their default values. The logistic

regression model has a fairly high accuracy. The most

logical explanation for this behavior may be the prob-

lem of the unbalanced dataset.

4.3.5 Multilayer Peceptron

Before training the MLPerceptron model, a method

was used to standardize observations with the Stan-

dardScaler integrated tool. This standardization was

performed on the train set and the test set with the

use of the ﬁt transform method of the StandardScaler

class. The process of standardizing the train and test

set is done because the MultiLayer Perceptron algo-

rithm is sensitive to feature scaling. After standard-

ization, you create an instance of the imported model.

For the MLP model there are many parameters that

can be changed such as the number of hidden lay-

ers(hidden layer sizes), the choice of activation func-

tion for hidden layers (activation), the optimization of

weights between nodes (solver), and the number of

eras (max iter). In this research work, only the num-

ber of eras has changed, i.e. max iter, which is tradi-

tionally a large number, often hundreds or thousands,

and allows the algorithm to run until the model error

has been sufﬁciently reduced to the minimum.

5 RESULTS AND DISCUSSION

Different algorithms were available for ML. This

project used linear models and only scratched the

surface of deep learning with multi-layer perceptron.

During background study, were found to have used

deep learning models. Deep learning tends to per-

form better than linear models. This project should

have included deep learning methods than linear mod-

els. The accuracy score achieved by optimizing the

models is satisfactory. In KNN model with the op-

timization of the hyperparameter k you can see how

all the evaluation metrics have improved.The logis-

tic regression model has a fairly high accuracy. The

most logical explanation for this behavior may be the

problem of the unbalanced dataset. As can be seen

from the result the Random Forest model is able to

classify all outcomes very well excluding the diabetes

category lada presumably because it is the category

with the least samples.Finally seeing the results of

MLP model you can see that in support of an accu-

AI4EIoTs 2020 - Special Session on Artiﬁcial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives

538

Table 2: Type 2 diabetes.

Model/Metrics Precision Recall F1-Score

KNN 0.90 0.91 0.90

KNN(k=1) 0.99 0.95 0.97

LR 0.93 0.99 0.96

DT(optimized) 1.00 0.97 0.98

RF 0.99 0.99 0.99

MLP 1.00 0.99 0.99

Table 3: Type 1 diabetes.

Model/Metrics Precision Recall F1-Score

KNN 0.31 0.19 0.24

kNN(k=1) 0.76 0.92 0.83

LR 0.93 0.70 0.80

DT(optimized) 0.85 1.00 0.92

RF 0.97 0.98 0.98

MLP 0.94 1.00 0.97

Table 4: Gestational diabetes.

Model/Metrics Precision Recall F1-Score

KNN 0.37 0.80 0.49

KNN(k=1) 0.83 1.00 0.91

LR 0.20 0.07 0.10

DT(optimized) 0.94 1.00 0.97

RF 0.94 1.00 0.97

MLP 0.94 1.00 0.97

racy of 98.9% there are also very high precision and

recall values.This data makes us understand that for

the dataset used to classify the type of diabetes the

MultiLayer Perceptron model is the best in terms of

rating metrics, but perhaps not in terms of the speed

of execution of the algorithm, which contrary to the

metrics takes a signiﬁcant time to run the algorithm.

Below will be the results of the classiﬁers developed

for each type of diabetes. The tables show the results

of precision, recall and f1-score in a way One vs All,

that is, for each individual type of diabetes are con-

sidered positive samples belonging to that type and

all other negatives. Accuracy was not included in the

result comparison tables because it would lead to a

foreclosing due to the imbalance of the dataset.

Table 5: MODY diabetes.

Model/Metrics Precision Recall F1-Score

KNN 0.68 0.91 0.79

KNN(k=1) 0.89 1.00 0.94

LR 0.78 0.58 0.67

DT(optimized) 0.89 1.00 0.84

RF 0.96 1.00 0.98

MLP 0.96 1.00 0.98

Table 6: LADA diabetes.

Model/Metrics Precision Recall F1-Score

KNN 0.32 0.55 0.40

KNN(k=1) 0.79 1.0 0.88

LR 0 0 0

DT(optimized) 0.73 1.00 0.85

RF 0.88 0.64 0.74

MLP 0.85 1.00 0.92

6 CONCLUSIONS AND FUTURE

WORKS

The idea underlying this work was to investigate and

build, through ML techniques, classiﬁcation models

that can accurately predict the type of diabetes. The

search for a better prediction and therefore a faster di-

agnosis by the specialized medical staff, can be vital

for the health of the diabetic patient, forced to face

numerous complications that can become lethal if not

managed in a timely manner. To complete this re-

search some further work has to be done, since the

system is already in a prototyping stage and it’s not

ready yet to be made available to the public, since it

lacks in some aspects concerning the safety and vali-

dation of the models performing the prediction. Any-

way, the results we obtained can be considered satis-

factory and motivate us to go further in our investiga-

tions beyond completing the robustness and safety as-

pects of the prototyping system. Future developments

of this research carried out include:

• Reﬁne the techniques of extracting features from

raw data, in particular, integrate data also from

CGM, insulin pump, Artiﬁcial pancreas System

like OpenAPS and platform like Tidepool;

• Solve the problem of dataset imbalance with un-

dersampling and/or oversampling algorithms for

more accurate classiﬁcation of classes with fewer

samples;

Machine Learning Approaches for Diabetes Classiﬁcation: Perspectives to Artiﬁcial Intelligence Methods Updating

539

• Broaden the work done for the classiﬁcation of

type of diabetes in other medical areas as well;

• Consider using a NO-SQL database(like Mon-

goDb) to manage the many data and features made

available from an electronic medical record;

ACKNOWLEDGEMENTS

The research presented in this paper was co-funded

by the activities of Gesan s.r.l. within the Tabl-

Health Project (CUP: B49J17000710008). This re-

search was also partially funded by the project ”At-

trazione e Mobilit

a dei Ricercatori” Italian PON Pro-

gramme (PON AIM 2018 num. AIM1878214-2).

REFERENCES

Alssema, M., Vistisen, D., Heymans, M., Nijpels, G.,

umer, C., Zimmet, P. Z., Shaw, J. E., Eliasson, M.,

Stehouwer, C. D., Tab

ak, A. G., et al. (2011). The

evaluation of screening and early detection strategies

for type 2 diabetes and impaired glucose tolerance

(detect-2) update of the ﬁnnish diabetes risk score for

prediction of incident type 2 diabetes. Diabetologia,

54(5):1004–1012.

Bennetts, C. J., Owings, T. M., Erdemir, A., Botek, G.,

and Cavanagh, P. R. (2013). Clustering and classiﬁ-

cation of regional peak plantar pressures of diabetic

feet. Journal of biomechanics, 46(1):19–25.

Brunese, L., Mercaldo, F., Reginelli, A., and Santone, A.

(2019). Neural networks for lung cancer detection

through radiomic features. In 2019 International Joint

Conference on Neural Networks (IJCNN), pages 1–10.

IEEE.

Brunese, L., Mercaldo, F., Reginelli, A., and Santone, A.

(2020). An ensemble learning approach for brain can-

cer detection exploiting radiomic features. Computer

methods and programs in biomedicine, 185:105134.

Contreras, I. and Veh

ı, J. (2018). Artiﬁcial intelligence for

diabetes management and decision support: Literature

review. Journal of Medical Internet Research, 20.

Coronato, A. and Naeem, M. (2019). A reinforcement

learning based intelligent system for the healthcare

treatment assistance of patients with disabilities. In

International Symposium on Pervasive Systems, Algo-

rithms and Networks, pages 15–28. Springer.

Dankwa-Mullan, I., Rivo, M., Sepulveda, M., Park, Y.,

Snowdon, J., and Rhee, K. (2019). Transforming dia-

betes care through artiﬁcial intelligence: the future is

here. Population health management, 22(3):229–242.

Fiorini, N., Lipman, D. J., and Lu, Z. (2017). Cutting edge:

towards pubmed 2.0. Elife, 6:e28801.

Hossin, M. and Sulaiman, M. (2015). A review on evalua-

tion metrics for data classiﬁcation evaluations. Inter-

national Journal of Data Mining & Knowledge Man-

agement Process, 5(2):1.

IDF, I. (2017). International diabetes federation. idf dia-

betes atlas, 8th edn. brussels, belgium: International

diabetes federation, 2017.

Mercaldo, F., Nardone, V., and Santone, A. (2017). Dia-

betes mellitus affected patients classiﬁcation and diag-

nosis through machine learning techniques. Procedia

computer science, 112:2519–2528.

Meyfroidt, G., G

uiza, F., Ramon, J., and Bruynooghe, M.

(2009). Machine learning techniques to examine large

patient databases. Best Practice & Research Clinical

Anaesthesiology, 23(1):127–143.

Nanda, S., Savvidou, M., Syngelaki, A., Akolekar, R., and

Nicolaides, K. H. (2011). Prediction of gestational

diabetes mellitus by maternal factors and biomarkers

at 11 to 13 weeks. Prenatal diagnosis, 31(2):135–141.

Osman, A. H. and Aljahdali, H. M. (2017). Diabetes disease

diagnosis method based on feature extraction using k-

svm. Int J Adv Comput Sci Appl, 8(1).

Patterson, C. C., Karuranga, S., Salpea, P., Saeedi, P.,

Dahlquist, G., Soltesz, G., and Ogle, G. D. (2019).

Worldwide estimates of incidence, prevalence and

mortality of type 1 diabetes in children and adoles-

cents: Results from the international diabetes feder-

ation diabetes atlas. Diabetes Research and Clinical

Practice, 157:107842.

Ricciardi, C., Cantoni, V., Improta, G., Iuppariello, L.,

Latessa, I., Cesarelli, M., Triassi, M., and Cuocolo, A.

(2020a). Application of data mining in a cohort of ital-

ian subjects undergoing myocardial perfusion imaging

at an academic medical center. Computer Methods

and Programs in Biomedicine, page 105343.

Ricciardi, C., Edmunds, K. J., Recenti, M., Sigurdsson, S.,

Gudnason, V., Carraro, U., and Gargiulo, P. (2020b).

Assessing cardiovascular risks from a mid-thigh ct im-

age: a tree-based machine learning approach using

radiodensitometric distributions. Scientiﬁc Reports,

10(1):1–13.

Romeo, V., Cuocolo, R., Ricciardi, C., Ugga, L., Cocozza,

S., Verde, F., Stanzione, A., Napolitano, V., Russo, D.,

Improta, G., et al. (2020). Prediction of tumor grade

and nodal status in oropharyngeal and oral cavity

squamous-cell carcinoma using a radiomic approach.

Anticancer Research, 40(1):271–280.

Sahin, H. and Subasi, A. (2015). Classiﬁcation of the car-

diotocogram data for anticipation of fetal risks using

machine learning techniques. Applied Soft Comput-

ing, 33:231–238.

Verde, L. and De Pietro, G. (2018). A machine learning ap-

proach for carotid diseases using heart rate variability

features. In HEALTHINF, pages 658–664.

Verde, L., De Pietro, G., and Sannino, G. (2018). Voice dis-

order identiﬁcation by using machine learning tech-

niques. IEEE Access, 6:16246–16255.

Zheng, T., Xie, W., Xu, L., He, X., Zhang, Y., You, M.,

Yang, G., and Chen, Y. (2017). A machine learning-

based framework to identify type 2 diabetes through

electronic health records. International journal of

medical informatics, 97:120–127.

Zhou, H. and Myrzashova, R. (2020). Ai based systems for

diabetes treatment: a brief overview of the past and

plans for the future. Journal of Physics: Conference

Series, 1453:012063.

AI4EIoTs 2020 - Special Session on Artiﬁcial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives

540