Healthcare Analytics 2 (2022) 100118

Contents lists available at ScienceDirect

Healthcare Analytics
journal homepage: www.elsevier.com/locate/health

An assessment of machine learning models and algorithms for early

prediction and diagnosis of diabetes using health indicators✩
Victor Chang a ,∗, Meghana Ashok Ganatra b , Karl Hall b , Lewis Golightly a , Qianwen Ariel Xu a
a Department of Operations and Information Management, Aston Business School, Aston University, Birmingham, UK
b
School of Computing, Engineering and Digital Technologies, Teesside University, Middlesbrough, UK

ARTICLE INFO ABSTRACT

Keywords: Breakthroughs in healthcare analytics can help both the doctor and the patient. Analytics in healthcare can
Diabetes help spot and diagnose diseases early on. Therefore, they can also be used to improve healthcare quality and
Diabetic analysis patient outcomes. Machine learning models can be used to find patterns in data and generate predictions
Machine learning
based on these patterns. They are employed in healthcare applications for disease diagnosis, prognosis, and
Healthcare analytics
treatment. With the development of new algorithms and other technological innovations, these models have
become more effective than ever at delivering patient treatment. The primary objective of this research is to
apply different machine learning algorithms to predict the diagnosis of diabetes. Furthermore, these models
are compared to determine the most effective model in this regard by evaluating their accuracy of prediction,
alongside other performance metrics such as precision, recall and F1 score. Of the models investigated, Random
Forest significantly outperformed the others, achieving an accuracy of 82.26%.

1. Introduction 2045. In the US, spending on diabetes care and management makes up
roughly 25% of all healthcare spending [2]. This also does not consider
As one of the most common diseases in the world, diabetes mellitus the indirect costs of diabetes, such as productivity loss. The cost of
affects 37.3 million Americans in 2019 or 11.3% of the country’s popu- diabetes is expected to increase further due to an estimated increase in
lation. The proportion of diabetes-related deaths in the US is estimated the prevalence of diabetes from 10.5% in 2021 to over 12% in 2045.
to be between 11.5% and 12%, and significantly higher among obese In addition, the cost of diabetes per individual in North America was
people at 19.4% [1]. There are two main types of diabetes, Type 1 around $8650 in 2021. From this, it can be posited that, as a whole,
and Type 2, both of which share many similarities and differences people’s health is worsening as a general observation, resulting in
in their underlying causes and recommended management strategies.
more people being affected by diabetes. Diabetes was diagnosed in 537
While accounting for around 8% of diabetes patients, Type 1 is much
million individuals globally in 2021; by 2045, that figure is projected to
more uncommon and is primarily caused by genetics, although not the
rise to 783 million. [3]. This forecast can also be partly attributed to the
only factor. While taking this into account, Type 2 diabetes receives
rate at which the global population is increasing. Additionally, it is also
more attention due to its prevalence. The exact biological mechanisms
estimated that half of all people affected by diabetes are undiagnosed.
that allow for Type 2 diabetes to manifest are more unclear, but it
is the case that genetic, environmental and lifestyle variables are all Diabetes is usually diagnosed in one of two ways – in the more
contributing factors in a variety of ways. Diabetes is not fully curable traditional way involving manual diagnosis by health practitioners
at this time, but medications such as insulin can be used to manage the – or by technology. Each of these methods has distinct advantages
symptoms carefully. The effective management of diabetes is crucial and disadvantages. While it is true that manual diagnosis by health
to prevent additional complications, such as eye, foot, and mouth practitioners allows for human expert insight, advances in technology
problems, kidney disease and even certain cancers. Like other diseases, have made this approach much more effective as time goes on and is
one of the best ways to effectively manage diabetes lies in diagnosing becoming the preferred approach. Another advantage of technological
the disease early before the effects become more serious. approaches is that they require less time and resources to employ. Addi-
Diabetes also has a significant economic impact, with estimated tionally, in the initial stages of the disease, indicators of diabetes can be
annual expenditures for direct costs relating to diabetes increasing from easier to identify through technology than by manual examinations [4]
$966 billion in 2021 to a projected value in excess of over $1 trillion in while eliminating human errors and complications in the initial analysis

✩ This work is partly supported by VC Research (VCR 0000191) for Prof Chang.
∗ Corresponding author.
E-mail addresses: [email protected], [email protected] (V. Chang).

https://doi.org/10.1016/j.health.2022.100118
Received 7 September 2022; Received in revised form 8 October 2022; Accepted 15 October 2022

2772-4425/© 2022 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

stages. As the availability of electronic health record data continues positive results, with all models with SVM-linear providing the best
to increase, it becomes more attractive to utilize automated diabetes accuracy (89%) and precision (88%). In addition, KNN provided the
diagnosis systems. In particular, artificial intelligence (AI) and machine best recall (90%) and F1 score (88%). Overall, the work suggests that
learning (ML) approaches are the two main ways automated diabetes SVM-linear and KNN are the optimal algorithms for the diabetes dataset
diagnosis systems can be built. and are the best for a diabetes diagnosis.
Lu et al. [6] highlight the use of ML with a network approach for
1.1. Aims and objectives disease prediction of Type 2 diabetes. They used patient records from
private health insurance formulated into a graph (bipartite) projected
Broadly speaking, the main goal of this study is to develop an to the patient network. They used specific features and characteristics
automated diabetes diagnosis system by utilizing a suite of ML models. to train eight ML algorithms for the ability to predict. The experiments
The key objectives are outlined as follows: showed positive effects with AUC (ranging from 0.79–0.91) when the
experiments were performed. The research results convey that a combi-
• Identify the most significant risk factors for Type 2 diabetes nation between ML techniques and network analysis can be positively
and the correlation between them to apply ML techniques more used for successful disease risk prediction in the diabetes domain.
effectively for diabetes diagnosis. Nadeem et al. [7] convey a fusion-based ML approach to predicting
• Identify and analyze the diabetes data to improve its suitability onset diabetes. The architecture performs data fusion to prepare a
by implementing sampling techniques and analysis. coherent dataset from various locations to be better aligned to the ML
• Compare the effectiveness of a range of different ML models by algorithms — which are then developed through the fusion of two
looking at several evaluation metrics to determine the best ML well-known ML algorithms that are used, which are SVM and ANN. The
solution for diabetes diagnosis. results demonstrate a classification accuracy of 94.67%, which exceeds
• Identify limitations with automated diabetes diagnosis systems the performance of other ML models for diabetes prediction by ∼1.8%.
and discuss how advances in this area can be explored to advance However, the diagnosis and classification of severe medical conditions
research in this area further are challenging because of low volume and low-quality contextual data
for the training and validation of algorithms which can compromise the
1.2. Research contributions results.
Sarwar et al. [8] demonstrate six well-known ML algorithms in the
• The relationships between all dependent variables are explored healthcare domain for diabetes prediction, namely SVM, KNN, Logistic
to discover the most important contributing factors to a person Regression (LR), Decision Trees (DT), Random Forest (RF) and Naïve
developing diabetes. It was discovered that the main health in- Bayes (NB). The predictions were made on the PIMA Indian dataset
dicators are Body Mass Index (BMI), age, blood pressure levels, containing 768 records. In the experiments, it has been observed that
cholesterol levels, general health, physical health, walking dif- SVM and KNN give the highest accuracy in predicting the conditions
ficulties, and income are the main contributing risk factors for showing 77% accuracy. The research can be expanded to a larger
diabetes. dataset to aim for an accuracy of 99.99%.
• Five machine learning classifiers were utilized to predict type 2 Sneha and Gangil [9] display an analysis of diabetes for early
diabetes: Decision Tree, Logistic Regression, Random Forest, K- condition prediction. The research proposes the design of a prediction
Nearest Neighbor, and Gaussian Naive Bayes classifiers. Univari- algorithm using ML algorithms. The method proposed aims to focus
able and multivariable attribute analysis was used to investigate on selecting attributes that aid in the early detection of diabetes using
the associations of potential risk factors with type 2 diabetes. predictive analysis. The research results show that the DT and RF al-
Principal Component Analysis (PCA) was applied to reduce the gorithms have the highest specificity of 98.20% and 98%, respectively,
dimensionality of the dataset. Synthetic Minority Oversampling and are the best for analyzing diabetic data.
Techniques (SMOTE) were used to stabilize the imbalance in the Hasan et al. [10] investigate diabetes prediction using an array of
output variable. ML classifiers. This technology is particularly challenging when there
• People with low social incomes are more likely to be burdened is a limited number of labeled data and missing values in the diabetes
by illness-related treatment because income and diabetes have a datasets. The algorithms used are KNN, DT, RF, AdaBoost, NB, XGBoost
strong link. This is especially true in nations like the US, where and Multilayer Perceptron. The experiments in the research were all
treatment expenses are expensive. A more thorough study should conducted using the Pima Indian diabetes dataset. The proposed clas-
be done in the future to address the rising cost diabetes has on sifier is the best performing with sensitivity, specificity, false omission
the economy to reduce the financial burden of diagnostic and rate, diagnostic odds ratio, and ROC-AUC. Furthermore, the classifier
nondiagnostic costs. was shown at the time to outperform the state-of-the-art by 2%.
• A thorough analysis of how to prevent death due to this condition, Sisodia, D. and Sisodia, DS [11] use classification algorithms to
further information such as electrocardiogram results for diag- predict diabetes early on. Three ML algorithms are evaluated on var-
nosed individuals, carnitine levels, and glucose level parameters ious methods performing experiments on the Pima Indians diabetes
should be investigated. dataset. The results showed an accuracy reading of 76.30% using
an NB classification algorithm. The research can be expanded in the
2. Related literature future by improving the automation of diabetes analysis using other
ML algorithms.
The literature reviewed in this section focuses on studies employing Rajesh and Sangeetha [12] use a range of ML classifiers for predict-
multiple ML algorithms to achieve the diagnosis of diabetes in the ing diabetes diagnosis, including NB, KNN, SVM, and a range of DT
initial stages of the disease. algorithms such as ID3 and C4.5. They conducted their research using
Kaur and Kumari [5] explore predictive modeling and analytics for the Pima Indians dataset. They also used an RF algorithm to obtain a
diabetes using ML. They utilized five different models to detect the result of 100% accuracy, but they determined that since their RF model
condition: Support Vector Machine (SVM) with the linear kernel, SVM suffered from data overfitting, the results should not be acknowledged.
with the RBF kernel, K-Nearest Neighbor (KNN), Artificial Neural Net- Instead, they achieved an accuracy of 90.62% by utilizing the C4.5 DT
works (ANN), and Multifactor Dimensionality Reduction algorithms. algorithm. Since this algorithm is frequently used and widely successful
Each model was evaluated based on parameters such as accuracy, in medical applications, they concluded this model to be the best for
recall, precision, F1 score, and ROC-AUC. The experiments highlighted diabetes diagnosis systems.

2
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Kelarev et al. [13] focused on the application of DT and Ensemble accuracy readings of 94.87%, demonstrating higher than other systems
classifiers to predict cardiovascular autonomic neuropathy in diabetes and saving many lives.
patients. Ensemble classifiers refer to hybrid ML models which take Goyal and Jain [21] proposed a novel approach by integrating a
from more than one model in order to improve their performance. number of classification algorithms, such as NB, LR, J48 DT, SVM
In their research, they used ensemble models based on DT, namely and RF using 10-fold cross-validation techniques and ensemble Bagging
the ADTree, J48, NBTree, RandomTree, REPTree and SimpleCart DTs. algorithms. Of these, the highest accuracy was demonstrated by SVM
Additionally, they investigated how various ensemble techniques could as a single classifier at 77.34%, whereas the accuracy of the ensemble
be optimized in such a way. Their best results were achieved by LR-based model was even higher at 77.60%. Results of the experiment
applying AdaBoost to ensemble Bagging and DT models, achieving an show a good form of increment in the accuracy of the classifier with
accuracy of 94.84%. low error rate and enhanced ROC-AUC when applied with 10-fold cross-
Ganie and Malik [14] also employed Ensemble classifiers for early- validation. This type of approach is very helpful in medical diagnosis
stage diabetes detection. They collected the dataset from different and can enable the medical practitioners to take appropriate decisions
departments of hospitals of Jammu and Kashmir-UT, India, such as for chronic diseases, including diabetes.
inpatient, outpatient, and emergency. By evaluating the effectiveness of Abdulhadi and Al-Mousa [22] explored the use of supervised learn-
three ensemble learning techniques, i.e., Voting, Boosting, and Bagging ing models that could help assist doctors in the early detection of
to several classical machine learning algorithms, this study found that diabetes to improve the quality of patient’s lives. Their paper presented
Bagged Decision Tree outperformed with an accuracy rate of 99.14%. multiple techniques that were used to train multiple models. The mod-
Han et al. [15] studied the application of a multitude of different ML els they utilized were LR, Linear Discriminant Analysis, Linear SVM,
models for diabetes diagnosis prediction, including traditional models Polynomial SVM, RF and Voting Classifier. Of these, RF achieved the
such as SVM, RF and DT, alongside ensemble models focused on using highest accuracy of 82%.
rule extractions from SVM. They found that by combining this approach Gupta et al. [23] proposed two prediction models through the
with a RF classifier, the precision scores improved when compared to use of quantum ML and deep learning techniques. The aim of their
the base RF and SVM models (89.6% compared to 81.2% and 88.4%, study was to assess and compare the predictive capabilities of these
respectively). Conversely, they found the recall scores of the ensemble models when compared to other state-of-the-art models using the Pima
classifier (44.3%) were lower than the base RF model (49.0%) but Indian Diabetes Dataset. Of the two models proposed, the deep learning
still higher than the base SVM model (40.0%). This was attributed to multilayer perceptron model proved to be the most effective predictor
the SVM + RF ensemble using simpler rules than the base RF model. with 95% accuracy when using four hidden layers, and outperformed
Nevertheless, they concluded that using the SVM + RF ensemble was all other DL models by at least 7.36%. The quantum ML model returned
still preferred over either of the base models. an accuracy of 86% at best.
In the work of Hassan et al. [16], they employed unsupervised meth- Sivaranjani et al. [24] applied ML algorithms after applying dimen-
ods for building a diabetes detection model. The model employed the sionality reduction and feature selection to predict diabetes using the
K-means clustering technique to group the features and the Silhouette Pima Indian Diabetes Dataset. More specifically, their study used step
and Elbow methods were used to determine the optimal number of forward feature selection and step backward feature elimination and
clusters. Then, they applied several algorithms (Multilayer Perceptron, principal component analysis and their application to SVM and RF
RF, DT, SVM, and KNN) to the created cluster-based dataset and models. The results of each classifier is evaluated using 5-fold cross
complete dataset, the evaluation results show that RF achieved the best validation. RF produced similar predictive performance regardless of
accuracy of 99.57% on the cluster-based dataset. whether step forward or step backward feature selection was employed,
Ramesh et al. [17] created a healthcare monitoring framework to at 82.9%. SVM performed slightly worse, but the use of step backward
manage diabetes remotely. The framework uses ML algorithms em- feature selection proved more fruitful than step forward, increasing its
bedded in real-time diabetes prediction and the interconnection of performance by 1.5% up to 81.41%.
various devices (physician portals, smart patient devices) to enable Finally, Lama et al. [25] study the use of ML methodologies with
healthcare experts to make informed decisions remotely, while enabling the Stockholm Diabetes Preventive Program to investigate the most
cost reduction and closed-loop communication. important risk factors for developing Type 2 diabetes using SHAP de-
Khanam and Foo [18] compared various ML techniques for diabetes pendence values. They discovered that the most significant contributors
prediction using the Pima Indian Diabetes Dataset. Throughout the towards developing Type 2 diabetes were BMI, waist-hip ratio, age,
study, they used seven ML algorithms (DT, KNN, RF, NB, AB, LR, SVM), blood pressure and genetics. Furthermore, they showed that having
with all algorithms achieving greater than 70% accuracy. Their study a combination of these risk factors increased the risk of developing
showed KNN and AB algorithms to achieve the best results for diabetes diabetes multiplicatively.
prediction, with both achieving accuracy scores of 79.42%. KNN had To find the most accurate model to predict diabetes, Refat et al. [26]
slightly higher precision and F1-measure scores than AB, so it can be applied six machine learning models, namely XGBoost, RF, DT, SVM,
considered slightly better in this regard. LR, and KNN. They also used deep learning base classification tech-
Krishnamoorthi et al. [19] created a framework for disease predic- niques, including LSTM, ANN, and MLP. They discovered that has a
tion in healthcare by utilizing ML techniques. The study enforced ML 100% train and test accuracy of zero train and test loss, and one for pre-
techniques on the Pima Indian Diabetes Dataset. The results demon- cision, recall, F1-score, and ROC-AUC accuracy, with an execution time
strated that the LR performed better than the other ML algorithms. of 46.074 s, making it the most accurate model for this classification.
For diabetes disease, it has been highlighted that there is a correlation Although the RF test accuracy is 92.3%, it should be emphasized that it
between glucose and BMI. One limitation of the study is that they used is still effective. However, compared to ML models in this study, deep
a structured dataset and aimed to test with an unstructured dataset. It learning models performed worse. An ANN provided good performance
is suggested that the methods used in the paper can be easily applied for three different deep learning models with accuracy ranging from
to other healthcare domains to predict other diseases. 88.5% to 92.3%.
Ahmed et al. [20] used decision-level Fusion ML as a decision
support system for predicting diabetes. It has been noticed that many 3. Data and methodology
ML techniques have been used for the prediction, but the authors
aimed to focus on the accuracy element with the various proposed The methodology of the study has six key steps (Fig. 1), consisting
models. For this study, two particular ML techniques have been used, of a literature survey, data acquisition, exploratory analysis, data pre-
namely SVM and ANN. As a result, the fuzzy detection system presented processing, model selection and model evaluation. Firstly, existing

3
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Table 1
Comparison of accuracy results from a literature survey.
Study Model(s) Accuracy (%)
Kaur and Kumani [5] SVM-Linear 89
Nadeem et al. [7] SVM-ANN ensemble 94.67
Sarwar et al. [8] SVM and KNN 77
Sissodia and Sissodia [11] NB 76.30
Rajesh and Sangeetha [12] C4.5 DT 90.62
Kelarev et al. [13] Adaboost DT with Bagging 94.84
Ganie and Malik [14] Bagged DT 99.14
Hassan et al. [16] K-means clustering and RF 99.57
Khanam and Foo [18] KNN and AB 79.42
Ahmed et al. [20] Fusion ML Decision 94.87
Goyal and Jain [21] LR ensemble 77.60
Abdulhadi and Al-Mousa [22] RF 82
Gupta et al. [23] Multilayer Perceptron 95
Sivaranjani et al. [24] SVM with feed backward feature elimination 82.9

Fig. 1. Methodology of the study.

diabetes-related research was surveyed to develop the research ques- data scientists to spot anomalies, discover trends, check assumptions,
tions and select an appropriate solution. Identifying a suitable dataset and derive better insights for their tasks [27].
is important for constructing an effective classifier. Therefore, to accu- In our study, a correlation matrix is first computed to understand
rately predict diabetes, the dataset was obtained from BRFSS, which the relationship between each pair of variables. If there is a strong
provided sufficient data samples and attributes for training purposes. relationship between attributes, removal of the attributes should be
In the exploratory analysis, the dataset was cleaned to ensure that considered. Through several types of data visualization techniques, the
there were no missing values, and then exploratory data analysis was distribution of the attributes and the data balance are analyzed to
conducted to get a greater understanding of the data for pre-processing. improve understanding of the dataset and its features.
DT, RF, KNN, NB and LR were selected as the algorithms to build Feature engineering helps identify the features that hold the most
the classifiers and the performance of the classifiers was analyzed and relevant information to the predicted target [28]. PCA is used as the
compared for accuracy, sensitivity, specificity, ROC-AUC curve, and feature selection tool to reduce the dimensionality of data and optimize
Precision–Recall Curve. the model. Moreover, it can preserve as much information present in
the complete data as possible [29].
3.1. Data description
3.3. Data pre-processing
The dataset used in this study is a subset of Behavioral Risk Factor
Surveillance System (BRFSS) data. It is a clean dataset of 70,692 Before feeding the data to the model, it must be partitioned into
responses to the CDC’s BRFSS2015 survey. Participants without dia- train and test sets. The training partition is used to train the model,
betes, those with prediabetes, and those with diabetes are split equally. and the test partition is used to assess its performance. The provided
This dataset has 253,680 records and twenty-one unbalanced feature dataset is split in an 80:20 ratio for training and testing, respectively.
variables. A more detailed description of these variables is shown The choice of this ratio follows the Pareto principle, the main idea of
in Appendix. Diabetes_binary is the two-class binary target variable. which is that 80% of the effect in most cases comes from 20% of the
A value of 0 indicates that the patient does not have diabetes, while a causes [30]. When the dataset was cleaned, 5% of null values were
score of 1 indicates that they have prediabetes or diabetes.1 added by using sample functions to all of the variables except for the
target variable.
3.2. Exploratory data analysis Training the model with this missing data might result in inaccurate
performance. Therefore, data entries containing missing values were
Exploratory Data Analysis is a method used to investigate and removed from the dataset for better results. The MinMax Scaler was
analyze data using various visual techniques. These techniques allow applied to normalize the data values before fitting the dataset. The
given dataset is extremely biased, with just 15% of positive samples and
1
https://www.kaggle.com/datasets/alexteboul/diabetes-health-indicators- the rest being negatives. To address this problem, SMOTE techniques
dataset were applied to oversample entries from the minority class, resulting

4
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

in an equal distribution dataset with over 200,000 data entries. As the

data is categorial, and the target variable is binary in nature, we apply
classification models to determine the accuracy of prediction.

3.3.1. SMOTE
SMOTE [31] is a statistical sampling technique used to increase the
number of examples in the dataset uniformly. The component creates
new instances based on minority cases. It takes the entire dataset as an
input, but it increases the percentage of only the minority cases. The
proportion of the majority of cases is unaffected by the implementation
of SMOTE.
The new occurrences are different from minority cases that already
exist. Instead, each target class and its close neighbors are sampled
from the feature space by the algorithm. The algorithm then creates
fresh examples incorporating traits from both the target case and
its neighbors. With this method, each class has access to additional
characteristics, and the samples are more inclusive.

3.3.2. ADASYN Fig. 2. Bar plot showing diabetes status and general health.
In order to produce more synthetic data for minority class examples
that are more challenging to learn than minority class examples that are
easier to learn, ADASYN can be used to utilize weighted distribution for 3.4. Machine learning classifiers and evaluation metrics
separate minority class instances. The ADASYN technique thereby en-
hances learning by reducing the bias brought on by the class imbalance Five different ML algorithms were applied to build the classifiers
and adaptively shifting the classification decision boundary toward the using the training set. A summary of the algorithms used is summarized
challenging samples. in Table 2:
The dataset used to train the ML classifiers was imbalanced, with A number of algorithms have been used for the prediction of dia-
80% of the cases having a target variable value of 0 and 20% of the betes, and we have identified the most popular of these and selected
cases having a target variable value of 1. Such a large data imbalance these five algorithms to predict our data for the following reasons:
could potentially result in a high likelihood of inaccurate results, and
• When using XGBoost, optimal performance is achieved on
we need only to increase the percentage of minority class to balance
datasets with at least eight variables. Therefore, it was determined
instances in the target variables evenly. Therefore, SMOTE techniques
not to be appropriate for this dataset.
are shown to be more suitable when compared to ADASYN for this
• SVM is not ideal for this dataset because it has slow calculation
dataset.
times when processing large datasets.
• Linear Regression and KNN regressor were not selected as they
3.3.3. Principal component analysis are regression models and not suitable for this study.
Principal component analysis (PCA) [32] is an unsupervised sta- • ANN and other related deep learning techniques were not chosen
tistical method that reduces the dataset dimensions. The use of PCA as they would potentially be overfitting the model.
facilitates linking variables by identifying relationships between them. • After analyzing all these scenarios, the following five classifiers
It thereby simplifies the data dimensionality but also retains the impor- were selected: DT, RF, KNN, LR and NB.
tant existing trends and patterns within the data.
After training the classifier, four results are obtained by comparing
the output with the predetermined labels: true positive, true negative,
3.3.4. T-distributed stochastic neighbor embedding
false positive and false negative, which constitute the confusion ma-
T-distributed stochastic neighbor embedding [33] (t-SNE) converts
trix. Evaluation metrics based on the confusion matrix are used to
a high dimensional dataset into a low dimensional graph while retain-
verify the correctness of the ML model, including Accuracy, Sensitivity,
ing a substantial portion of the original data. It does this by giving each
Specificity, ROC-AUC curve, and Precision–Recall curve.
data point a specific place on a two- or three-dimensional map. By lo-
The pseudocode for the ML classifiers is outlined in Table 3.
cating clusters in the data, this approach makes sure that an embedding
preserves the meaning in the data. While reducing dimensionality, t-
4. Implementation and results
SNE seeks to keep comparable examples near together and dissimilar
instances away. T-SNE is non-linear, so it can capture the structure of 4.1. Exploratory data analysis
trickier manifolds and involves hyperparameters, unlike PCA.
Based on the above features of both PCA and t-SNE, it is clear that General health is measured on a numerical scale of 1 to 5. From
PCA is a more suitable approach as only the variables that do not Fig. 2, we can determine that those with a general health score of 3–5
contribute much to the target variable should be removed. There is are more prone to develop diabetes. In this chart, class 0 indicates no
no such requirement for keeping comparable samples used in t-SNE diabetes and 1 indicates prediabetes or diabetes. There are over 80,000
together in the dataset. patients with a general health of 3, 20,000 with a general health of
Datasets with high dimensionality can lead to overfitting when 4, and approximately 10,000 with a general health of 5 for class 0
utilizing ML classifiers. The dataset used for this study consists of (no diabetes), indicating that many people who are not diabetic should
twenty-one variables, meaning that it can be considered to have high improve their overall health to avoid contracting the disease since they
dimensionality. After applying PCA, eight variables, namely BMI, age, are at high risk by having a general health score in this range. For the
high blood pressure, high cholesterol, general health, physical health, more than 20,000 patients with class 1 diabetes or prediabetes, they
walking difficulty and income [34], were identified, contributing the must improve their physical health with medication under the guidance
most to the prediction of a diabetes diagnosis. of a doctor to improve their life quality and overall physical health. The

5
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Table 2
ML classifiers.
ML classifier Summary
DT DT is a tree-structured classifier wherein the split at each node is chosen while the internal and leaf node samples are
kept to a minimum. A maximum depth of 50 was chosen and criterion entropy was used to evaluate the leaf
cleanliness.
RF RF builds decision trees out of several samples and classifies them based on the majority vote from each tree. As the
data was clean and binary, we applied ten trees as depth. In order to acquire the performance, criteria entropy and a
random state of 42 were used.
KNN The nearest neighbors are found using a mix of Ball Tree (BT), KD Tree (KDT), or brute-force search. The Manhattan
or Euclidean distance functions are used for computing the classifications. A k-value of 3 is used as it was determined
to be the best k-value for the dataset.
LR LR is most commonly used for the classification of binary classes. It internally uses the sigmoid function to learn the
linear relationships between variables. The solvers and penalty can be changed to see visible differences in the
classification. In this study, a random state of 0 was used.
NB Naive Bayes is a probabilistic AI calculation based on the Bayes Theorem that is used in a wide range of classification
problems.

Table 3
Pseudocode for early prediction of diabetes using ML classifiers.
1 Import dataset and data
2 Dimensionality Reduction: PCA
3 Class Imbalance: SMOTE
4 Identify the ML classifiers that will be used for the classification
5 MLC=[ DecisionTreeClassifier ( ), RandomForestClassifier ( ), KNeighborsClassifier ( ),
LogisticRegression ( ), NB ( )]
6 for (i=0;i<5;i++) do
model = MLC[i]
model.fit ( );
model.predict ( );
print(classification_report ( ), accuracy_score ( ), precision_score ( ), recall_score ( ),
f1_score ( ), hamming_loss ( ), roc_curve ( ), roc_auc_score( ))
end

government creates several health programs for them to enroll in and points are more predominant and continuous, it is also clear that pa-
improve their lifestyles by changing their diet and exercising apart from tients who have higher physical activity are less likely to have diabetes
taking their medication. than patients with low activity.
The distribution of patients’ health states compared to their diabetes In Fig. 7, the BMI distribution is the most frequent between 22 and
status is measured on a scale of 1 to 5, with 1 being the best and 5 being 30. BMI is one of the most crucial risk factors for developing diabetes.
the worst. Furthermore, we can see that the most significant difference Higher BMI correlates strongly with the probability of occurrence of
in the form of the distribution is that the density of class 0 is more on a the disease.
scale of 1 to 3. Still, the density of class 1 is more widely dispersed from The correlation plot shown in Fig. 8 shows the correlation between
1 to 5, indicating that the health state of diabetic patients is generally all the features and there is no linear correlation between feature
well managed. variables.
The overall form and distribution of the recommendations for edu- The data is equally sampled using SMOTE for class 0 and class 1 in
cation compared to the diabetes status of patients (quartiles relatively the output target variable to get more efficient predictions when ML
close to each other) are similar in the next example. However, there algorithms are applied with less biased results (see Fig. 9).
are more outliers in the case of the positive class. As shown in Fig. 10, the components below 0.4 are not contributing
much towards the target variable and will make little to no difference
In a comparison of patient income when compared to diabetes
in the output, so that they can be dropped.
status, we can see that the density of the plot increases with income
The SMOTE technique is applied to overcome the imbalance in the
level, indicating that the higher the income, the higher the chances of
output variable. PCA is applied to reduce the number of features and
getting the disease. In contrast, people with class 1 diabetes have a plot
select only those that contribute the most to the target variable, leaving
evenly distributed on a scale of 1 to 8, indicating that all income level
eight components (see Fig. 11).
groups have the disease, while it is denser among the income levels of
5 to 8.
4.2. Machine learning models
In Fig. 4, it can be seen that 96.5% of patients had high cholesterol
levels, and 94.7% are not heavy drinkers. These are both considered 4.2.1. Logistic regression
severe risk factors for developing diabetes. Regression analysis is a predictive modeling analysis since it always
Diabetes has a moderate risk of stroke and heart disease, according deals with predictions. Regression is divided into three categories. The
to this study. The majority of patients did not have a stroke or heart first one is linear regression. One of the important tools used in the
disease, as seen in Fig. 5. Furthermore, diabetic patients had a frac- natural sciences is logistic regression, which is strongly associated with
tional value of an extra number of strokes or heart attacks compared to neural networks in sociology. It simply classifies and determines if a
healthy people. Even healthy people can have a stroke or heart attack, specific event is occurring or not, which addresses many categorization
but their risk is lower than that of individuals who have already been problems. The output of LR predicts the outcome in binary form for a
diagnosed with diabetes. discrete dependent variable.
From Fig. 6, it is shown that patients with no diabetes have lesser It uses discrete dependent variables and is a statistical classification
mental health issues when compared to patients with diabetes. While model. The logistic function or sigmoid function, which accepts a value
diabetes patients have slightly high mental health issues as their data between 0 and 1, is how LR operates. The phrase ‘‘logistic’’ derives from

6
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 3. Violin plots showing the diabetes status of patients and income.

Fig. 4. Pie charts for cholesterol levels and heavy alcohol consumption.

the algorithm’s primary function, the Logit function. It is a controlled 4.2.2. Support Vector Machine (SVM)
learning algorithm that deals with likelihood and only allows for two SVM is known to perform well in many areas, like bioinformat-
alternative outcomes: yes or no, true or false, 0 or 1, high or low. The ics, content, picture acknowledgment, speaker ID, pattern recognition
in photographs, and target recognition. This categorization technique
probability is rounded to 0 if it is less than 0.5 and to 1 if it is more than
works with both linear and non-linear data. It performs classification
0.5. The S-shaped sigmoid function presented below takes real values by building a hyperplane in high- or infinite-dimensional space that,
and plots them in the [0,1] range, as shown in (1). under ideal circumstances, isolates the data into two classes that can
1 be used for classification or regression, as shown in Fig. 12. For a given
cos (𝑦) = (1) set of objects, several different separation hyperplanes are feasible.
(1 + exp (−𝑦))

7
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 5. Bar charts for diabetes status with respect to stroke and heart disease.

Fig. 6. Box plots for diabetes status with respect to mental health and physical activity.

4.2.3. K-Nearest Neighbor (KNN)

KNN [36] is a classification algorithm that presents a method to
manage data requests that assess how likely it is that data is thought
to be a person from a social event based on several data restrictions
currently in place. This tactic is known as a sluggish methodology
since it would not build a structure utilizing the train set unless an
instructional assortment request was made. It also goes by the name
‘‘non-parametric approach’’. This indicates that since the technique is
based on data, it makes no assumptions about hidden data scattering.
Red triangles, blue squares, and green circles are the datapoints that
represent the various classes, as shown in Fig. 13.

4.2.4. Random Forest

RF is defined as a process that creates numerous decision trees
by referring to each tree to make decisions. Typically, n number of
datapoints are picked from the dataset, and by combining them, a
stable decision is produced. If there are more guesses, the average
of all predictions is used. The classification and regression problems
Fig. 7. BMI distribution among patients. are resolved using the RF [37] technique. As depicted in Fig. 14, a
considerable number of trees are produced in the forest throughout this
process. More trees mean a healthier forest, which produces findings
The hyperplanes should not be located close to the datapoints but with great accuracy.
should be selected, so they are far from the information focal points The RF architecture is made up of several trees. Every tree offers
for each categorization. Support vectors are used to describe datapoints a particular selection. By averaging all of the options, the most recent
that are more closely spaced from the hyperplane [35]. prediction is evaluated.

8
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 8. Correlation plot of data features.

Fig. 9. Count plot for the target variable.

Fig. 10. Cumulative explained variance compared to the number of components.

4.2.5. Naive Bayesian

Naive Bayesian is a term used to describe a grouping algorithm, • P(a/z) represents the target/object class’s back likelihood for the
a probabilistic classifier that relies on the autonomous supposition provided predictor.
between several indicators. The method that uses the dataset as its
source of data conducts research and predicts the class grade using ◦ The target/object class’s earlier likelihood is P(a).
Bayes’ Theorem. It determines whether there is a chance that the ◦ The probability, or P(a/c), measures how likely a predictor
input data will be classified and helps predict the class of the test for is given a class.
obscure information. This style of organizing works well with large
• The earlier predictor likelihood is denoted by P(z).
datasets. Using the Bayes Theorem equation shown below, it is possible
to calculate the back likelihood for each class.
( ( ) )
( ) 𝑃 𝑎𝑧 ∗ 𝑃 (𝑎) 4.2.6. Decision Trees
𝑎
𝑃 = (2) Decision Trees are used to extract data from a vast number of
𝑧 𝑃 (𝑧)
( ) ( ( ) ( )) ( ) available datasets using decision rules. Data that can be quickly stored
𝑎 𝑧1 𝑧2 𝑧𝑛
𝑃 = 𝑃 ∗𝑃 ∗…∗𝑃 ∗ 𝑃 (𝑐) (3) and further classified using a decision tree is simply categorized. In this
𝑧 𝑎 𝑎 𝑎

9
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 13. KNN representation for k = 3.

Fig. 11. Dimensionality reduction of diabetes dataset.

Fig. 14. Random Forest.

After training, the classifiers are evaluated using several metrics

Fig. 12. Representation of support vectors in the SVM model. based on the confusion matrix (Fig. 16). In this matrix, TP (true
positive) denotes the number of samples for which the model correctly
predicted the positive category. TN (true negative) is the number of
study, we discuss various decision-tree-based techniques for classifying samples for which the model correctly predicted the negative category.
data (see Fig. 15). FP (false positive) refers to the number of samples for which the model
For example, according to the DT above, smokers tend to pass away incorrectly predicted the positive category. FN (false negative) is the
early. If a person does not smoke, whether or not they drink is the next number of samples in the negative category that the model incorrectly
consideration. A person becomes old and dies if they do not drink or predicts.
smoke. The metrics include accuracy, sensitivity, specificity, ROC-AUC
A person dies elderly if they drink alcohol, do not smoke, and weigh curve, and Precision–Recall curve. Confusion matrices and the formulae
less than 90 kg. If a person drinks but does not smoke, their weight is used to calculate these metrics are defined below.
considered. Finally, if a person does not smoke, does drink, and weighs Accuracy computes the number of correct predictions a model
more than 90 kg, they are more likely to die early. makes, and hamming loss is the number of incorrect predictions:
𝑇𝑃 + 𝑇𝑁
Accuracy = (4)
4.3. Evaluation of results 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
In comparison to all other models for Diabetes (Table 4), RF has the
Before feeding the data to the model, the dataset is cleaned, and highest accuracy (82.26%), which means that RF was 82.26% correct
the data values are normalized before fitting the dataset. SMOTE tech- in distinguishing cases with and without diabetes in the test sample.
niques are used to oversample entries from the minority class, resulting NB performed the worst, with an accuracy rate of 70.56%. However,
in an equal distribution dataset with over 200,000 rows. The dataset is accuracy is not an adequate performance measure for assessing the
split in an 80:20 ratio in accordance with the Pareto principle. power of the whole model on its own and is not truly predictive.

10
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 15. Decision Tree flowchart.

Fig. 17. Random Forest confusion matrix.

Fig. 16. Confusion matrix. i.e., recall, is 80.45% for RF, meaning 80.45% of the cases with diabetes
in the test set were correctly selected by the RF classifier. NB is the
worst performer in this perspective, with a recall rate of 67.07%. False
Precision, Recall, and Specificity are used alongside accuracy to provide negatives should be avoided, as missing the presence of the disease can
a more balanced evaluation approach. lead to treatment being delayed and can cause real damage [38].
𝑇𝑁
Hamming loss = 1 − Accuracy (5) Specificity = (8)
𝐹𝑃 + 𝑇𝑁
The Hamming loss is the proportion of incorrect labels to total Specificity measures the percentage of actual negatives that were
labels. Hammering loss is determined in multi-class classification as accurately identified. This is accomplished by dividing the number of
the hamming distance between ‘actual’ and ’predictions.’ Hamming loss correctly predicted negative samples by the total number of samples
penalizes only the individual labels in multi-label categorization. The that were either correctly or mistakenly forecasted as negative (TN,
lower the loss, the better the model performance. RF has the lowest FP). The specificity of DT and RF reached over 84% with 84.05% and
Hamming loss of 17.74%. 84.07%, respectively, indicating that they can correctly identify over
𝑇𝑃 84% of the cases without diabetes in the test set. A detection system
Precision = (6) with a high specificity contributes to the issue of over-medicalization,
𝑇𝑃 + 𝐹𝑃
In disease detection, predicted false positives can lead to misdi- as diagnosing a patient without diabetes as a person with diabetes can
agnosis and cause wastage of healthcare resources, and improving cause anxiety and unnecessary follow-up procedures [39].
the precision of diagnostic models can help to improve this problem. 2 (𝑅𝑒𝑐𝑎𝑙𝑙 × 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
F − measure = (9)
Precision quantifies the number of correctly forecast positive observa- 𝑅𝑒𝑐𝑎𝑙𝑙 + 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
tions: This is accomplished by counting the samples that were correctly An effective diabetes detection system should avoid both missed and
predicted as positive (TP) and dividing them by the total number misdiagnosis, but accuracy and specificity are conflicting performance
of positive predictions, correct or incorrect (TP, FP). According to metrics. F-measure accounts for both precision and recall. It can have
Table 4, of all the classifiers evaluated, the RF model had the highest a maximum value of 1, signifying flawless precision and recall, and a
precision of 83.47%, followed by DT at 83.02%, indicating that they minimum value of 0 if either precision or recall is zero. RF has the
correctly predicted more than 83% of all cases predicted to be diabetic. highest F1 score of 82.26%, while it achieved the best performance
LR performed the worst in this perspective, with a Precision rate of in all other evaluation metrics, which validates it as the best diabetes
70.56%. classifier in this study.
𝑇𝑃 Table 4 and Fig. 18 shows the comparison of the performance
Recall = (7)
𝑇𝑃 + 𝐹𝑁 metrics of accuracy, precision, sensitivity, specificity, F1 score and
Recall or sensitivity, like precision, aims to figure out the proportion hamming loss. These metrics evaluate the classifiers’ performance [40,
of true positives that were accurately detected. It accomplishes this 41]. According to Table 4, RF was the best classifier for diabetes
by dividing the correctly predicted positive samples (TP) by the total detection on the test set in terms of every metric. Moreover, as shown
number of positives, either correctly or incorrectly predicted as positive in Fig. 17, RF was more accurate in predicting cases with diabetes
(TP, FN). Recall measures the number of correct positive predictions (Precision: 83%) than in predicting cases without diabetes (Precision:
out of all possible positive predictions made. The highest sensitivity, 81%), whereas it was less able to identify cases with diabetes from all

11
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Table 4
Performance metrics for the ML models.
Classifier Accuracy Precision Sensitivity Specificity F1-Score Hamming loss
DT 81.02% 83.02% 77.98% 84.05% 81.10% 18.98%
RF 82.26% 83.47% 80.45% 84.07% 82.26% 17.74%
KNN 80.55% 81.20% 79.50% 81.59% 80.54% 19.45%
LR 72.64% 72.06% 73.95% 71.34% 72.64% 27.36%
NB 70.56% 72.09% 67.07% 74.04% 70.52% 29.44%

in-depth information on the variables that can be helpful in predicting

the disease in a better way. In order to achieve this, cross-validation
techniques should be used to improve the performance of the model
and include SHAP values to understand which features have the highest
impact on the outcome variable. In order to compete with state-of-
the-art models achieving above 90% accuracy, the implementation of
ensemble models should be utilized.

5.2.1. Federated machine learning

Originally proposed by Google, the idea behind federated machine
learning (FML) is to construct ML models from data that is distributed
across several devices or servers to prevent data leakage [42]. This
allows for a more collaborative and decentralized approach while also
reducing distribution costs, data distribution imbalances and reliability.
While employing similar strategies to distributed machine learning,
Fig. 18. Visual comparison of ML model performance. FML places a heavier emphasis on data privacy and protection, partic-
ularly during the stage of training the model. Additionally, federated
database systems share a lot of similarities in methodology but put
positive samples (Recall: 80%) than it was to identify cases without more of an emphasis on database operations as opposed to protect-
diabetes (Recall: 84%). For the DT and KNN (Table 4), although DT ing data privacy. While FML is not associated with improving the
(Accuracy: 81.02%) was only 0.47% more accurate than KNN (Accu- performance of ML models, particular attention needs to be paid to
racy: 80.55%) when classifying all test samples, DT (Precision: 83.02%) approaches that strive to improve data privacy. In the age of Big Data
was 1.82% more accurate than KNN (Precision: 81.20%) in predicting and rapidly improving information systems, this is true more than ever,
cases with diabetes, and DT (Specificity: 84.05%) was 2.46% less likely and its importance will only continue to increase. Ultimately, utilizing
to misdiagnose diabetes than KNN (Specificity: 81.59%). In addition, FML massively increases data privacy. In projects similar to ours,
the scores of the hamming loss have slight variation, with KNN having whereby sensitive patient data is analyzed, FML should be employed
a higher hamming loss. A lesser value of hamming loss indicates a when appropriate to ease people’s concerns in this regard.
better classifier. Therefore, DT is a better classifier than KNN. KNN
(Sensitivity: 79.50%) had the advantage that it was 1.52% less likely to 5.2.2. ADASYN
miss a diabetes case than DT (Sensitivity: 77.98%). NB performed worst Adaptive synthetic sampling (ADASYN) is a sampling approach for
among the classifiers, and it may have the issue of missed diagnosis as learning from imbalanced datasets. Similar to SMOTE, the purpose of
its precision score is less than 70%. ADASYN is to address class imbalance within datasets to reduce bias
The Precision–Recall curve is perfect when it is right-angle or associated with imbalanced data and shift the classification decision
boundaries towards the more difficult data points. In order to achieve
perpendicular. The ROC-AUC curve and Precision–Recall curve for RF
this, data from the minority class is adaptively generated, which is
in Fig. 19 are highly perpendicular, showing the model to be a good
harder to learn. In the initial ADASYN proposal [43], performance
fit. The AUC score of the RF classifier is 0.8226, and its hamming loss
was compared against SMOTE across a standardized list of datasets.
of 0.1774 reaffirms that it is comparatively better than other classifiers
ADASYN was found to outperform SMOTE when the same ML models
to predict the disease.
were used after addressing the data imbalance with these two tech-
niques. Furthermore, it was noted that to improve performance further,
5. Conclusions
ADASYN should be integrated into ensemble ML classifiers by using
bootstrap sampling techniques and then embedding ADASYN into the
5.1. Research contributions
sampled data.

The key contribution of our research was the development of pre- 5.2.3. Ensemble classifiers
dictive models that used ML to detect people who were developing Ensemble ML classifiers are a more recent development and are
diabetes. This work presented a study of five classifiers (DT, RF, KNN, well-known for increased predictive performance compared to tradi-
LR and NB) for predicting the likelihood of diabetes. The RF classifier tional approaches. Ensemble classifiers combine algorithms from two or
achieved the highest accuracy of 82.26%. more ML models to outperform each of the individual models. The most
This research assessed the prediction of diabetes based on the common technique used in ensemble classifiers is boosting. The aim of
key features. With the enhanced capability of the ML algorithms in boosting is to improve the performance of weak learning algorithms,
classification, the model can significantly aid medical practitioners in such as classification rules or decision trees [44]. AdaBoost and its
the diagnosis. variants are a commonly used set of iterative boosting algorithms
which forces weak learning models to focus on the more complex
5.2. Limitations of the study and future work data points by performing more iterations and producing additional
classifiers [45]. Bootstrap aggregating, or bagging, is another com-
One future direction of this study is to identify the gene and clinical mon component of such classifiers, whereby composite classifiers are
variables that affect diabetes. A more suitable dataset can give more created by combining the outputs of the various models to produce

12
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

Fig. 19. ROC-AUC & Precision–Recall curve for RF classifier.

Table A.1
Description of dataset.
S. No Features Description
1 Diabetes binary 0 = no diabetes, 1 = diabetes
2 HighBP 0 = no high BP 1 = high BP
3 HighChol 0 = no high cholesterol 1 = high cholesterol
4 CholCheck 0 = no 1 = yes
5 BMI Body Mass Index
6 Smoker Have you smoked at least 100 cigarettes in your entire life? 0 = no 1 = yes
7 Stroke You had a stroke. 0 = no 1 = yes
8 HeartDiseaseorAttack Coronary Heart Disease (CHD) or myocardial infraction (ML) 0 = no 1 =yes
9 PhysActivity physical activity in past 30 days - not including job 0 = no 1 = yes
10 Fruits Consume Fruits 1 or more times per day 0 = no 1 = yes
11 Veggies Consume Vegetables 1 or more times per day 0 = no 1 = yes
12 HvyAlcoholConsump Heavy drinkers (adult men having more than 14 drinks per week and adult women having
more than 7 drinks per week) 0 = no 1 = yes
13 AnyHealthcare Have any kind of health care coverage, including health insurance, prepaid plans such as
HMO, etc. 0 = no 1 = yes
14 NoDocbcCost Was there a time in the past 12 months when you needed to see a doctor but could not
because of cost ? 0 = no 1= yes
15 GenHlth Would you say that in general your health is ?
Scale 1–5 1= excellent 2 = very good 3 = good 4 = fair 5 = poor
16 MentHlth Now thinking about your mental health, which includes stress, depression, and problems with
emotions, for how many days during the past 30 days was your mental health not good ? scale
1–30 days
17 PhysHlth Now thinking about your physical health, which includes physical illness and injury, for how
many days during the past 30 days was your physical health not good ? scale 1–30 days
18 DiffWalk Do you have serious difficulty walking or climbing stairs? 0 = no 1 = yes
19 Sex 0 = female 1 = male
20 Age 13-level age category (_AGEG5YR see codebook) 1 = 18–29 9 = 60–64 13=80 or older
21 Education Education level (EDUCA see codebook) scale 1–6 1 = Never attended school or only
kindergarten 2 = Grades through 8 (Elementary) 3 = Grades 9 through 11 (some high school)
4 = Grade 12 or GED (High school graduate) 5 = College 1 year to 3 years (Some college or
technical school) 6 = College 4 years or more (College graduate)
22 Income Income scale (INCOME2 see codebook) scale 1–8 1 = less than $10,000 5 = less than $35,000
8 = $75,000 or more

a more robust prediction. It can be hypothesized that implementing Data availability

ensemble classifiers and associated techniques would improve pre-
dictive performance, as other research extensively suggests this to The data that has been used is confidential.
be the case. In summary, to improve upon the results conducted in
this study, ADASYN-based ensemble classifiers can be developed, fine-
Acknowledgment
tuned, and optimized to build a more effective automated diabetes
diagnosis system.
This work is partly supported by VC Research, UK (VCR 0000191)
for Prof Chang.
Declaration of competing interest

The authors declare that they have no known competing finan- Appendix
cial interests or personal relationships that could have appeared to
influence the work reported in this paper. See Table A.1.

13
V. Chang, M.A. Ganatra, K. Hall et al. Healthcare Analytics 2 (2022) 100118

References [23] H. Gupta, H. Varshney, T.K. Sharma, et al., Comparative performance analysis of
quantum machine learning with deep learning for diabetes prediction, Complex
[1] A. Stokes, S.H. Preston, Deaths attributable to diabetes in the United States: Intell. Syst. 8 (2022) 3073–3087, http://dx.doi.org/10.1007/s40747-021-00398-
comparison of data sources and estimation approaches, PLoS One 12 (1) (2017) 7.
e0170219. [24] S. Sivaranjani, S. Ananya, J. Aravinth, R. Karthika, Diabetes prediction using ma-
[2] M.C. Riddle, W.H. Herman, The cost of diabetes care—an elephant in the room, chine learning algorithms with feature selection and dimensionality reduction, in:
Diabetes Care 41 (5) (2018) 929–932. 2021 7th International Conference on Advanced Computing and Communication
[3] J. Elflein, Estimated number of diabetics worldwide in 2021, 2030, and Systems, ICACCS, Vol. 1, IEEE, 2021, pp. 141–146.
2045, 2022, [online] Available at: https://www.statista.com/statistics/271442/ [25] L. Lama, O. Wilhelmsson, E. Norlander, L. Gustafsson, A. Lager, P. Tynelius,
number-of-diabetics-worldwide/ (accessed 27 Sep, 2022). et al., Machine learning for prediction of diabetes risk in middle-aged Swedish
[4] J. Chaki, S.T. Ganesh, S.K. Cidham, S.A. Theertan, Machine learning and artificial people, Heliyon 7 (7) (2021) e07419.
intelligence based diabetes mellitus detection and self-management: a systematic [26] M.A.R. Refat, M. Al Amin, C. Kaushal, M.N. Yeasmin, M.K. Islam, A comparative
review, J. King Saud Univ.-Comput. Inf. Sci. (2020). analysis of early stage diabetes prediction using machine learning and deep
[5] H. Kaur, V. Kumari, Predictive modelling and analytics for diabetes using a learning approach, in: 2021 6th International Conference on Signal Processing,
machine learning approach, ACI 18 (1/2) (2020) 90–100, http://dx.doi.org/10. Computing and Control, ISPCC, IEEE, 2021, pp. 654–659.
1016/j.aci.2018.12.004. [27] A.M. Malik, A.K. Sagar, S. Sahana, Prediction of cardiopathy using exploratory
[6] H. Lu, S. Uddin, F. Hajati, M.A. Moni, M. Khushi, A patient network-based data analysis, in: 2021 IEEE 6th International Conference on Computing,
machine learning model for disease prediction: The case of type 2 diabetes Communication and Automation, ICCCA, IEEE, 2021, pp. 117–122.
mellitus, Appl. Intell. 52 (3) (2022) 2411–2422, http://dx.doi.org/10.1007/ [28] A. Thakkar, R. Lohiya, Attack classification using feature selection techniques:
s10489-021-02533-w. a comparative study, J. Ambient Intell. Humaniz. Comput. 12 (1) (2021)
[7] M.W. Nadeem, H.G. Goh, V. Ponnusamy, I. Andonovic, M.A. Khan, M. Hus- 1249–1266.
sain, A fusion-based machine learning approach for the prediction of the [29] C.L. Chowdhary, D.P. Acharjya, Segmentation and feature extraction in medical
onset of diabetes, Healthcare 9 (10) (2021) 1393, http://dx.doi.org/10.3390/ imaging: a systematic review, Procedia Comput. Sci. 167 (2020) 26–36.
healthcare9101393. [30] H.B. Harvey, S.T. Sotardi, The pareto principle, J. Am. College Radiol. 15 (6)
[8] M.A. Sarwar, N. Kamal, W. Hamid, M.A. Shah, Prediction of diabetes using ma- (2018) 931.
chine learning algorithms in healthcare, in: 2018 24th International Conference [31] N.V. Chawla, K.W. Bowyer, L.O. Hall, W.P. Kegelmeyer, SMOTE: synthetic
on Automation and Computing (ICAC) Newcastle Upon Tyne, United Kingdom,
minority over-sampling technique, J. Artificial Intelligence Res. 16 (2002)
2018, pp. 1–6, http://dx.doi.org/10.23919/IConAC.2018.8748992.
321–357.
[9] N. Sneha, T. Gangil, Analysis of diabetes mellitus for early prediction using
[32] H. Abdi, L.J. Williams, Principal component analysis, Wiley Interdiscip. Rev.
optimal features selection, J. Big Data 6 (1) (2019) 13, http://dx.doi.org/10.
Comput. Stat. 2 (4) (2010) 433–459.
1186/s40537-019-0175-6.
[33] L. Van der Maaten, G. Hinton, Visualizing data using t-SNE, J. Mach. Learn. Res.
[10] M.K. Hasan, M.A. Alam, D. Das, E. Hossain, M. Hasan, Diabetes prediction
9 (11) (2008).
using ensembling of different machine learning classifiers, IEEE Access 8 (2020)
[34] T. Sharma, M. Shah, A comprehensive review of machine learning techniques
76516–76531.
on diabetes detection, Vis. Comput. Ind., Biomed. Art 4 (1) (2021) 30, http:
[11] D. Sisodia, D.S. Sisodia, Prediction of diabetes using classification algorithms,
//dx.doi.org/10.1186/s42492-021-00097-7.
Procedia Comput. Sci. 132 (2018) 1578–1585.
[35] World Health Organization, Diagnostic Criteria and Classification of Hypergly-
[12] K. Rajesh, V. Sangeetha, Application of data mining methods and techniques for
caemia First Detected in Pregnancy (No. WHO/NMH/MND/13.2), World Health
diabetes diagnosis, Int. J. Eng. Innov. Technol. (IJEIT) 2 (3) (2012).
[13] A.V. Kelarev, A. Stranieri, J.L. Yearwood, H.F. Jelinek, Empirical study of Organization, 2013.
decision trees and ensemble classifiers for monitoring of diabetes patients in [36] G.O. Campos, A. Zimek, J. Sander, R.J.G.B. Campello, B. Micenková, E. Schubert,
pervasive healthcare, in: 2012 15th International Conference on Network-Based I. Assent, M.E. Houle, On the evaluation of unsupervised outlier detection:
Information Systems, IEEE, 2012, pp. 441–446. measures, datasets, and an empirical study, Data Min. Knowl. Discov. 30 (4)
[14] S.M. Ganie, M.B. Malik, An ensemble machine learning approach for predicting (2016) 891–927, http://dx.doi.org/10.1007/s10618-015-0444-8.
type-II diabetes mellitus based on lifestyle indicators, Healthcare Anal. 2 (2022) [37] J.W. Lee, J.B. Lee, M. Park, S.H. Song, An extensive comparison of recent
100092. classification tools applied to microarray data, Comput. Statist. Data Anal. 48
[15] L. Han, S. Luo, J. Yu, L. Pan, S. Chen, Rule extraction from support vector (4) (2005) 869–885, http://dx.doi.org/10.1016/j.csda.2004.03.017.
machines using ensemble learning approach: an application for diagnosis of [38] L. Zwaan, H. Singh, The challenges in defining and measuring diagnostic error,
diabetes, IEEE J. Biomed. Health Inf. 19 (2) (2014) 728–734. Diagnosis 2 (2) (2015) 97–103.
[16] M.M. Hassan, S. Mollick, F. Yasmin, An unsupervised cluster-based feature [39] A. Swift, R. Heale, A. Twycross, What are sensitivity and specificity? Evidence-
grouping model for early diabetes detection, Healthcare Anal. (2022) 100112. Based Nursing 23 (1) (2020) 2–4.
[17] J. Ramesh, R. Aburukba, A. Sagahyroon, A remote healthcare monitoring [40] H. Lai, H. Huang, K. Keshavjee, A. Guergachi, X. Gao, Predictive models for
framework for diabetes prediction using machine learning, Healthcare Technol. diabetes mellitus using machine learning techniques, BMC Endocr. Disord. 19
Lett. 8 (3) (2021) 45–57. (1) (2019) 101, http://dx.doi.org/10.1186/s12902-019-0436-6.
[18] J.J. Khanam, S.Y. Foo, A comparison of machine learning algorithms for diabetes [41] L. Zhang, Y. Wang, M. Niu, C. Wang, Z. Wang, Machine learning for character-
prediction, ICT Express 7 (4) (2021) 432–439. izing risk of type 2 diabetes mellitus in a rural Chinese population: the Henan
[19] R. Krishnamoorthi, S. Joshi, H.Z. Almarzouki, P.K. Shukla, A. Rizwan, C. Kalpana, Rural Cohort Study, Sci. Rep. 10 (1) (2020) http://dx.doi.org/10.1038/s41598-
B. Tiwari, A novel diabetes healthcare disease prediction framework using 020-61123-x, Art. (1).
machine learning techniques, J. Healthcare Eng. (2022). [42] Q. Yang, Y. Liu, T. Chen, Y. Tong, Federated machine learning: Concept and
[20] U. Ahmed, G.F. Issa, M.A. Khan, S. Aftab, M.F. Khan, R.A. Said, T.M. Ghazal, applications, ACM Trans. Intell. Syst. Technol. 10 (2) (2019) 1–19.
M. Ahmad, Prediction of diabetes empowered with fused machine learning, IEEE [43] H. He, Y. Bai, E.A. Garcia, S. Li, ADASYN: Adaptive synthetic sampling approach
Access 10 (2022) 8529–8538. for imbalanced learning, in: 2008 IEEE International Joint Conference on Neural
[21] P. Goyal, S. Jain, Prediction of type-2 diabetes using classification and ensemble Networks (IEEE World Congress on Computational Intelligence), IEEE, 2008, pp.
method approach, in: 2022 International Mobile and Embedded Technology 1322–1328.
Conference (MECON), IEEE, 2022, pp. 658–665. [44] L. Rokach, Ensemble-based classifiers, Artif. Intell. Rev. 33 (1) (2010) 1–39.
[22] N. Abdulhadi, A. Al-Mousa, Diabetes detection using machine learning classifi- [45] A. Shahraki, M. Abbasi, Ø. Haugen, Boosting algorithms for network intrusion
cation methods, in: 2021 International Conference on Information Technology, detection: A comparative evaluation of real AdaBoost, gentle AdaBoost and
ICIT, IEEE, 2021, pp. 350–354. modest AdaBoost, Eng. Appl. Artif. Intell. 94 (2020) 103770.

14