Abstract

Many intelligent healthcare systems have been developed to diagnose human diseases such as breast cancer, hepatitis, diabetes and heart diseases. Diabetes is a lifelong chronic disease that occurs when the pancreas does not produce enough insulin (Type I diabetes mellitus), or when the body's produced insulin is unable to be utilised properly (Type II diabetes mellitus), Researches that are carried out on diabetes using data mining techniques were done to predict type II diabetes mellitus using different diabetes datasets by different researchers; Pima Indians Diabetes Dataset (PIDD) is used by the majority of the researchers. The dataset (PIDD) has eight (8) attributes which limits more exploration in the field of Machine Learning (ML) for diabetes prediction. Diabetes prediction is limited because of the few attributes available in the diabetes datasets used, and these attributes play important roles in predicting diabetes mellitus types, classes and risk factors whenever a diabetes patient is diagnosed. This paper provides a systematic review of diabetes mellitus datasets, identifying the strength and weakness of the 8 attributes described in the PIDD, which is used by the most of the researchers. Furthermore, this paper has identified the need of the potential researchers in the research community to address the gap by enhancing the existing diabetes dataset attributes with additional attributes, identify the attributes required for the prediction of glucose level, diabetes Types, diabetes classes, diabetes risk factors and to develop a Model that can be used for the prediction.

1. Introduction

Health informatics is a rapidly expanding field concerned with the application of computer science and information technology to medical and health data. With the ageing population in developing and developed countries, as well as the rising cost of healthcare, governments and large health organisations are becoming increasingly interested in the potentials of Health Informatics to save time, money, and human lives; however, as a relatively new field, Health Informatics does not yet have a universally accepted definition ¹. The applications of Healthcare Informatics in clinical care decision-making, is the design, development and deployment of machine using data mining techniques that can help healthcare professionals to make effective and efficient clinical decisions” ¹.

Data Mining technique is among the most versatile techniques that have received a warm response in private organisations, government, enterprises and healthcare, it is mainly used in hospital for big data analytics and interpreting to smoothening the workflow of hospital management by helping health personal to serve patients better, and Surgeons to gain an insights to carry out the operation accurately with a great precision ². Data Mining Technology is this technology combines multiple disciplines such as statistics, probability, machine learning, and artificial intelligence to look for patterns and trends in massive amounts of data by utilising sophisticated mathematical algorithms to segment the data and assess the likelihood of future events ^{3, 4}. Typically, the Data Mining process can detect the patient's disease with great precision by uncovering the hidden information contained in the medical data gathered; this process entails a number of processes of working via interactive and iterative data sequences for determining the primary symptoms of communicable disease/ non-communicable diseases and then treats the patient well ². Non-Communicable Diseases (NCDs) which include stroke, heart disease, cancer, chronic lung cancer and diabetes are responsible for almost 70% of the deaths worldwide in which diabetes is the most common disease among them and the number of patients suffering from diabetes has quadrupled since 1980 ⁵.

2. Diabetes

Diabetes is a lifelong illness that occurs as a result of lack of insulin hormone or ineffectiveness of insulin hormone. Nowadays, diabetes is becoming one of the most serious diseases, which its frequency keep on increasing in the world and varies from one community to another based on age, gender, race, dietary habits, genetic characteristics and environmental factors ^{3, 6}. Insulin as a hormone transports glucose to the bodycells from bloodstream to be used as energy, if this energy not efficiently consumed by the bodycells, the excess can cause major health issues ⁷. There are two main types of diabetes that is Type I and Type II. According to health experts, diabetes occurs when the human body's gland called the pancreas cannot produce enough insulin (Type I diabetes), and the produced insulin cannot be used by the cell of the body (Type 2 diabetes) ^{6, 8, 9}.

Type I diabetes was previously known as Insulin-Dependent Diabetes Mellitus. This Type of diabetes can be formed at any age but needs to be diagnosed for those that are below 20 years of age. This Type of diabetes is formed when insulin-producing cells or beta cells in the pancreas get destroyed ¹⁰. Type I diabetes is caused by a damage to pancreatic and insulin-producing beta cells at the end of an autoimmune process, or due to unfamiliar disorders ⁶. In general, 5-10% of diabetes cases in the community constitute cases of Type I diabetes ⁸. Patients with Type I have insulin deficiency; they must take insulin hormone from outside for life ^{6, 8, 9}.

Type II diabetes was known as Non-Insulin-Dependent Diabetes Mellitus as it was diagnosed in patients above 20 years of age ¹⁰; or is one of the most common metabolic illnesses and is characterised by a deficiency in the generation of insulin secretion via the pancreatic islet β-cells ¹¹. Type II diabetes is generally associated with obesity and physical immobility. At the basis of the disease, genetically predisposed individuals have lifestyle-related insulin resistance and decreased insulin secretion over time ⁶. More than 90% of diabetes cases diagnosed worldwide is type II diabetes ⁷.

3. Related Research Work

Breault ¹², in his research work, Pima Indian Diabetic Dataset (PIDD) was used to test the developed model using data mining algorithms (naïve Bayes classifier) to accurately predict diabetic patient status of been positive or negative, where out of 392 complete cases, guessing all are non-diabetic gives an accuracy of 65.1%. The main objective of the research conducted by Parthiban et al. ¹³, was to predict the chances of the diabetic patient getting heart disease, the researchers applied Naïve Bayes technique, which produces an optimal prediction model, they proposed a system, which predicts attributes such as age, sex, blood pressure and blood sugar and the chances of a diabetic patient getting a heart disease, the dataset used in Parthiban et al. ¹³ work was a diabetic clinical dataset collected from Chennai of about 500 patients. Padmaja ¹⁴ aimed at find out the characteristics that determine the presence of diabetes and to track the maximum number of women who have diabetes. The researcher used clustering and attributes oriented induction techniques to track the characteristics of the women who have from diabetes using PIDD. Rajesh and Sangeetha ¹⁵, applied data mining techniques to classify diabetes clinical data and predict the likelihood of a patient who has diabetes or not, using the Decision Tree Algorithm (DTA) and PIDD.

Rahim ¹⁶ developed a preliminary classification and screening system using SVM algorithm for diabetic retinopathy, with the use of “Eye Fungus Images” diabetes dataset by focusing on the detection of the earliest signs of diabetic retinopathy, which are the microaneurysms. Hina et al. ⁵ carried out a research, using PIDD dataset and preprocessed it to a more meaningful structure; the data mining tool opted for the research was WEKA. In the research, different classifying algorithms such as Naïve Bayes, Multi-Layer Perceptron, Decision Tree, ZeroR, Random Forest, and Regression were applied to predict the chances of a diabetic patient being positive or negative. Neilesh and Gandhi ¹⁷ used a new feature selection method along with SVM in developing a model, in their work, the feature selection method was considered one of the best methods to improve the prediction accuracy of diabetes patient for having positive or negative, Pima Indian Diabetes Dataset was taken from the UCI Repository was used to test and train the developed Model. Vijayan and Anjali ¹⁸ developed a decision support system that used the different base classifiers (SVM, NB, Decision Stump, DT and AdaBoost algorithms), PIDD was used to test and train the developed system. Miss and Megha ¹⁹, conducted research where Back Propagation Neural Network (BPNN), and Graphical User Interface (GUI) were built using MATLAB, Pima Indian Diabetes Dataset was used by the researchers to test their proposed methodology. Mohebbi et al. ²⁰ developed an approach using deep learning for detection of Type 2 diabetes patient status using Continue Glucose Monitoring (CGM) signals dataset collected from 9 patients. Machine Learning Techniques such as Logistic Regression, Multi-Layer Perceptron, and Convolutional Neural Network were applied in the research, the dataset was divided into training and testing dataset, with 1 to 6 patients CGM signals used as a training dataset, and 7 to 9 patients CGM signals were used as a test dataset.

Francesco et al., ²¹ used Pima Indian Diabetic Dataset and WEKA tool to predict diabetic patients status of being positive or negative using different machine learning techniques such as Decision Tree, JRip, Multilayer Perceptron, Random Forest, HoeffdingTree, and BayesNet. Maham et al., ²², in their research, Pima Indian Diabetes Dataset was used to test their developed Model for the classification of diabetic patients as positive or negative using Multilayer Perceptron technique. Wenqian et al., ²³ used Pima Indian Diabetes Dataset, K-means for data reduction and Decision Tree as a classifier to predict the status of diabetic patient. The graph-based approach was developed by Mangrulkar ²⁴ where retinal image are classify by dividing retinal vessels in to two types as arteries and veins. In his research, the use of retinal vessels extracted for vascular changes detection is the most important phase. The decision tree algorithm having a 10-fold cross-validation method and K-means algorithm that was developed using WEKA is used by the researcher. The patient’s retinal image was used to find artery vein ratio. Sidong et al., ²⁵ in their research, five different techniques of machine learning were used for diabetes diagnosis and preprocessing of data, and those techniques include DNN, Logistic Regression, Decision Tree, SVM, and Naïve Bayes, Pima Indian Diabetic Dataset was used to calculate the accuracy of cross-validation. A diabetes prediction model with dropout was developed by Ashiquzzaman ²⁶, for diabetes prediction using deep learning neural networks that had a fully connected layer plus dropout layers. Pima Indian Diabetic Dataset was used to train and test the proposed Model. Deepti and Dilip ²⁷ developed a model having three different machine learning algorithms, those machine learning algorithms include Decision Tree, SVM, and Naïve Bayes, for diabetes status prediction of the target class of 1 as positive and 0 as negative, Pima Indian Diabetic Dataset was used to train and test the proposed Model. A data mining techniques for the prediction of Type 2 diabetes mellitus was developed by Han et al., ²⁸ using Pima Indian Diabetes Dataset to test their proposed Model for predicting status of diabetic patients by reducing dataset complexity and analysing the medical implication of every attribute and their correlation with diabetes mellitus. Safial and Islam ²⁹ developed a model using deep neural network with five-fold cross-validation and ten-fold cross-validation to diagnose diabetes using Pima Indian Diabetes Dataset. Ayon and Islam ³⁰ developed a strategy for diagnosis of diabetes using a deep neural network by training its attributes in a five-fold and ten-fold cross-validation fashion, Pima Indian Diabetes (PID) dataset used in the research for the prediction of diabetes patient status. Naz and Ahuja ³¹ presents a methodology for diabetes prediction using PIMA dataset. The researchers use Decision Tree (DT), Naive Bayes (NB), Artificial Neural Network (ANN), and Deep Learning (DL) techniques for the prediction of diabetes patient status. Alshammari et al., ¹¹ created a machine learning model that is capable of predicting diabetes with high performance, the researchers used the BigML platform to train four machine learning algorithms, namely, Deepnet, Decision Tree, Ensemble and Logistic Regression using collected dataset from Ministry of National Guard Hospital Affairs (MNGHA) at Saudi Arabia from 2013 to 2015, and the dataset attribute is for tested adult patients for Hemoglobin A1c (HgbA1c). In their research, HgbA1c result is used to determine the patient’s diabetic status where patient is classified as diabetic if HgbA1c value is grater or equal to seven, and patient is classified as non-diabetic if HgbA1c value is less than seven.

Bhoia et al., ³² developed a diabetes prediction model of females in Pima Indians heritage, using a binary classification problem with Pima Indians Diabetic Dataset, supervised learning techniques such as Support Vector Machine (SVM), classification tree (CT), Naïve Bayes (NB), k-Nearest Neighbour (k-NN), AdaBoost (AB), Random Forest (RF), Logistic Regression (LR), and Neural Network (NN) have been used in the research. The researchers use k-fold cross-validation to carry out the process of training and testing. Islam et al., ³³ in their research, they use a type II diabetes dataset taken from Bangladesh Demographic and Health Survey, 2011 containing 1569 respondents where 127 respondents are diabetes. For diabetes risk factors prediction, Six ML-based supervised classifiers as support vector machine, random forest, linear discriminant analysis, logistic regression, k-nearest neighbourhood, bagged classification and regression tree (Bagged CART) have been adopted in the research. Manimaran &Vanitha ¹⁰ in their research work, Maldives Bureau of Statistics (MBS) dataset was used, the dataset was collected from various districts to predict diabetes Disease using Data Mining Classification Techniques such as Multilayer Perceptron, Bayesian networks, Decision tree, and Fuzzy Lattice Reasoning (FLR), the dataset contains 1024 complete instances with 26 Parameters, and the data was gathered from answers to Questionnaires given during the research work. The main objective of the questionnaire was to converse on a set of parameters for the diagnosis of diabetes risk factors in patients. Alpan & İlgi ³⁴; Chaves & Marques ³, in their work, propose a comparative analysis of data mining techniques for early diabetes diagnosis, that is to predict only diabetes risk factors based on the attributes available in the dataset, both researchers use a publicly accessible dataset of Sylhet Diabetes Hospital, Portugal containing 520 instances, 17 attributes. Alpan & İlgi ³⁴ used WEKA tool, Bayesian Network, Naïve Bayes, Random Tree, Random Forest, k-NN, SVM techniques. The Results Obtained by the researchers indicated that k-NN performed the highest accuracy with 98.07% and this algorithm is the best method to identify and classify diabetes diseases on the studies dataset., while Chaves & Marques ³ uses six classification algorithms, namely k-nearest neighbours (kNN), Naive Bayes, Random Forest, Neural Network, AdaBoost, and Support Vector Machine (SVM), the Results Obtained by the researchers indicated that Neural Networks is good for diabetes prediction as the Model presents a Specificity of 97.5%., an Area Under the Curve (AUC) of 98.3%, F1-Score, Sensitivity and Precision of 98.4% and accuracy of 98.1%.

This paper has reviewed 28 papers and identified that many researches were done using different diabetes datasets; and the majority of the researchers give more impasses on predicting type II diabetes mellitus, this happened because of the limitation of attributes available in the datasets used as described in Table 2 below.

The Table 2 above confirmed that the majority of the reviewed papers used the Pima Indian Diabetes Dataset (PIDD). Therefore, this paper will consider PIDD attributes and suggest how the attributes limitations can be improved so that diabetes prediction can be enhanced to predict not only diabetes status, but also to predict patient Type of diabetes, class of diabetes and associated risk factors. This PIDD is available online from the URL https://data.world/data-society/pima-indians-diabetes-database, where all patients are females and of at least 21 years old, with 768 instances, 8 attributes and 1 outcome ^{31, 32, 35}.

Table 3 above described the 8 attributes that are available in PIDD with their descriptions, as stated below:-

1) Age: This attribute shows patient’s number of years, in numeric, of each instance, and the range of it value is from 21 to 81, where the average age value is 33 in PIDD.

2) Pregnancies: This attribute shows number of times a patient gets pregnant, in numeric, of each instance, and the range of it value is from 0 to 17, where the average pregnant value is 4.

3) Glucose: This attribute shows the level plasma glucose concentration or an Oral Glucose Tolerance Test result in 2 hours, and the range of it value is from 0 to 199, where the average is 121.

4) Blood pressure: This attribute shows the level of diastolic blood pressure in mm Hg, and the range of it value is from 0 to 122, where the average value is 69.

5) Skin thickness: This attribute shows the triceps skin thickness in mm, and the range of it value is from 0 to 99, where the average value is 21.

6) Insulin: This attribute shows the level of insulin, in numeric, and the range of it value is from 0 to 846, where the average value is 80.

7) BMI: This attribute shows body mass index in Kg/m² and the range of it value is from 0 to 67.1, where the average value is 32.

8) Diabetes pedigree function: This attribute shows the function scores of the likelihood of diabetes by inheritance, and the range of it value is from 0.078 to 2.42, where the range value is 0.47.

9) Outcome: This attribute is a class attribute, it shows the outcome of the prediction using either 0 and 1 or Yes and No, where 1 or Yes indicates diabetes patient is positive, while 0 0r No indicates the diabetes patient is negative.

The PIDD described all the 8 attributes presented in the dataset. However, there exist some strengths and weaknesses associated with the dataset. Table 4 below shows the PIDD strengths and weaknesses.

Table 4 provides and proves that:

1. Only one glucose attribute (Oral Glucose Tolerance Test result) is available in the PIDD, where the patient result of Oral Glucose Tolerance Test result will take minimum of 2 hours, therefore, there is a delay in predicting diabetic patient status.

2. Only patient age is captured in Age attribute, therefore, patient’s years with diabetes need to be captured as additional attribute to further predict diabetes types and classes.

3. The PIDD is limited to only 4 risk factors attributes namely: BMI, Diabetes Pedigree Function, Diastolic Blood Pressure and Skin Thickness. Where there are many more important risk factors such as obesity that is not available in PIDD.

4. The PIDD is limited to only one class attribute, therefore no class attribute as outcome of predicting risk factors.

5. The Pregnancy attribute is clearly limiting the PIDD of having male instance.

4. Conclusion

There are many issues bedeviling research in the field of diabetes using Machine Learning. Part of the issues related to diabetes datasets is the lack of enough attributes or even the diabetes datasets that are available online for free download. This issue limits the way diabetes related researches progress in the field of Machine Learning. There exist some researches that played the use of the Pima Indian Diabetes Dataset (PIDD) where Type II diabetes are mostly considered using the available attributes presented in the dataset. Type I diabetes, classes of diabetes mellitus and diabetes risk factors as well requires additional experimentation in the field of Machine Learning. Applying machine learning techniques to medical dataset is a trending research area, because there are too many healthcare diseases that require urgent result in investigation with accurate and efficient prediction ⁴. As it is discussed in the literature review, the majority of the diabetes datasets have only one class attribute. Therefore, the outcome of the prediction is always one. And for the non-class attributes, the attributes are limited to predicting only the status of the diabetic patient and no provision for predicting diabetes types, classes and risk factors. This paper identified some gaps that require attention from the research community. The identified gaps are:

1. To enhance the existing diabetes dataset attributes by providing the required additional attributes for diabetes prediction.

2. To identify attributes required for glucose prediction, diabetes types, classes prediction, and risk factors prediction.

3. To use a machine learning techniques to develop a model that can predict diabetes status, types of diabetes, classes of diabetes and diabetes risk factors.

References