applied

sciences
Article
Cardiovascular Health Management in Diabetic Patients with
Machine-Learning-Driven Predictions and Interventions
Rejath Jose † , Faiz Syed † , Anvin Thomas and Milan Toma *

College of Osteopathic Medicine, New York Institute of Technology, Old Westbury, NY 11568, USA;
[email protected] (R.J.); [email protected] (F.S.); [email protected] (A.T.)
* Correspondence: [email protected]
† These authors contributed equally to this work.

Abstract: The advancement of machine learning in healthcare offers significant potential for enhanc-
ing disease prediction and management. This study harnesses the PyCaret library—a Python-based
machine learning toolkit—to construct and refine predictive models for diagnosing diabetes mellitus
and forecasting hospital readmission rates. By analyzing a rich dataset featuring a variety of clini-
cal and demographic variables, we endeavored to identify patients at heightened risk for diabetes
complications leading to readmissions. Our methodology incorporates an evaluation of numerous
machine learning algorithms, emphasizing their predictive accuracy and generalizability to improve
patient care. We scrutinized the predictive strength of each model concerning crucial metrics like
accuracy, precision, recall, and the area under the curve, underlining the imperative to eliminate false
diagnostics in the field. Special attention is given to the use of the light gradient boosting machine
classifier among other advanced modeling techniques, which emerge as particularly effective in terms
of the Kappa statistic and Matthews correlation coefficient, suggesting robustness in prediction. The
paper discusses the implications of diabetes management, underscoring interventions like lifestyle
changes and pharmacological treatments to avert long-term complications. Through exploring the
intersection of machine learning and health informatics, the study reveals pivotal insights into al-
gorithmic predictions of diabetes readmission. It also emphasizes the necessity for further research
and development to fully incorporate machine learning into modern diabetes care to prompt timely
Citation: Jose, R.; Syed, F.; Thomas, interventions and achieve better overall health outcomes. The outcome of this research is a testament
A.; Toma, M. Cardiovascular Health to the transformative impact of automated machine learning in the realm of healthcare analytics.
Management in Diabetic Patients with
Machine-Learning-Driven Predictions Keywords: PyCaret; machine learning; stroke diagnosis; diagnostic accuracy; automated machine
and Interventions. Appl. Sci. 2024, 14,
learning; health informatics
2132. https://doi.org/10.3390/
app14052132

Academic Editors: Renáta Szabó and

Denise Börzsei 1. Introduction
Received: 5 February 2024 Diabetes mellitus is a condition in which the body’s metabolism of glucose is altered,
Revised: 26 February 2024 resulting in elevated blood glucose levels, termed hyperglycemia [1]. The major types of
Accepted: 28 February 2024 diabetes mellitus are related to the body’s production of or response to the hormone insulin,
Published: 4 March 2024 which decreases blood glucose levels through different mechanisms. The vast majority of
cases are either classified as diabetes mellitus type I, a condition with strong genetic com-
ponents in which there is an absolute deficiency of insulin production, or diabetes mellitus
type II, a condition with decreased sensitivity of insulin receptors to the hormone along
Copyright: © 2024 by the authors.
with instances of decreased production [2]. The chronic hyperglycemia that results from
Licensee MDPI, Basel, Switzerland.
diabetes leads to devastating long-term complications in patients such as the dysfunction
This article is an open access article
and possible failure of organs such as the kidneys, eyes, heart, and blood vessels, amongst
distributed under the terms and
others. Additionally, diabetes mellitus is a cerebrovascular and cardiovascular risk factor
conditions of the Creative Commons
Attribution (CC BY) license (https://
that increases the predisposition of a person to suffer an ischemic attack [3–5]. Hence,
creativecommons.org/licenses/by/
people with diabetes, particularly type 2 diabetes, are at an increased risk of developing car-
4.0/).
diovascular disease (CVD), such as heart attacks, strokes, and heart failure [6–8]. Diabetes

Appl. Sci. 2024, 14, 2132. https://doi.org/10.3390/app14052132 https://www.mdpi.com/journal/applsci

Appl. Sci. 2024, 14, 2132 2 of 23

can damage blood vessels and nerves that control the heart and blood vessels, leading to
heart disease over time [9–11]. High blood glucose levels in diabetes can cause damage
to blood vessels, which can result in serious heart complications [7,8]. It is important
for individuals with diabetes to manage their condition effectively to reduce the risk of
developing heart disease or stroke. Managing diabetes through proper glucose control,
regular exercise, a healthy diet, and other lifestyle modifications can help lower the chances
of developing CVD [7,8].
The management of diabetes mellitus involves a multidisciplinary approach in which
both lifestyle and pharmacologic interventions play a role. The goals of management are to
achieve normoglycemia and minimize cardiovascular and other long-term complications of
the disease [12]. Studies following patients with intensive glycemic control found decreased
incidences of major cardiovascular events in such patients [13]. In addition to pharma-
cologic management, patient education about the disease and long-term complications
has been shown in studies to improve glycemic control in patients [14–16]. The role of
behavioral modification on the effects of glycemic control has extensively been studied.
Interventions such as fasting, low-calorie diets, exercise, and behavioral psychological
therapy have all shown benefits in lowering blood glucose levels in studies assessing their
effects in patients with diabetes [17–21].
One high-priority healthcare quality measure is hospital readmission rates. With many
long-term complications, the burden of diabetes amongst hospitalized patients continues
to grow. Studies have correlated higher rates of readmission for hospitalized patients
with diabetes [22]. Adult patients with diabetes mellitus represent a sizable portion of
30-day unplanned hospital readmissions [23]. Methods that can be implemented to reduce
readmission rates include post-discharge coordination of care and support for patients with
diabetes, regular follow-ups with providers, specialty-specific care, and patient education.
Assessing readmission rates for patients with diabetes could benefit from additional studies
studying the effects of such interventions.
Machine learning has been explored as a tool to predict the incidence and risk factors
of diabetes and other cardiovascular diseases [24]. Previous studies utilizing logistical re-
gression and classification trees have predicted that fasting blood glucose, BMI, and age are
the main predictors of developing diabetes and observed population health risk assessment
for diabetes onset [25–27]. It has also been used to predict associated complications [28,29].
However, there remains a gap in the literature when it comes to analyzing outcomes and
readmission with the use of machine and deep learning.
The present study conducted a comparative analysis of machine learning techniques
to predict diabetes outcomes and readmission rates. The study used a dataset from the
Kaggle repository [30]. The models were deployed in PyCaret 3.0.3 to predict diabetes and
readmission rates. PyCaret has been previously utilized to predict the presence and severity
of diabetes [31], but this approach can be applied to study associated outcomes. The study
aimed to study machine learning methodologies while investigating their strengths and
weaknesses pertaining to diabetes prediction and analysis of readmission rates. By leverag-
ing large datasets and advanced algorithms, machine learning models can analyze complex
patterns and risk factors to accurately diagnose diabetes. This can lead to more-timely
interventions and improved patient outcomes. However, further research is needed to fully
establish machine learning as a cornerstone of modern healthcare diagnostics for diabetes.
Nevertheless, this study introduces an analytical framework that enhances the interpretabil-
ity and clinical applicability of predictive modeling for diabetes and its readmission risk,
using advanced machine learning techniques to not only analyze, but also translate the
public Kaggle dataset’s insights into actionable healthcare strategies, a critical advance-
ment beyond the mere performance evaluation of algorithms. The study uncovers several
insights into algorithmic predictions of diabetes readmission by exploring the intersection
of machine learning and health informatics. Ultimately, this study highlights the need
for additional research and development to fully integrate machine learning into modern
diabetes care, enabling timely interventions and improved overall health outcomes.
Appl. Sci. 2024, 14, 2132 3 of 23

The structure of the paper is laid out as follows: Section 2, “Materials and Methods”,
details the data acquisition process from the public Kaggle dataset and the methodologies
we applied in developing our analytical framework. Section 3, “Results”, unpacks the
performance metrics of our predictive models by delineating the rates of accurate classifica-
tions and misclassifications, as well as discussing the implications these have for clinical
practice and patient outcomes. It also explores the model’s capacity for generalization and
its robustness, ensuring a comprehensive understanding of its practical utility. In Section 4,
“Discussion”, we contextualize our findings within the broader landscape of healthcare
analytics and their implications for diabetes management.

2. Materials and Methods

This section describes the particular methods used for acquiring and preparing the
data, along with a detailed explanation of the machine learning pipeline utilized in the re-
search. Comparable methodologies may be found in other studies within this field [32–34].

2.1. Data Collection and Processing

In this study on diabetes and hospital readmission, we utilized a comprehensive open-
source dataset that includes clinical and demographic information. This dataset was used
to train and test machine learning models. The first dataset used in this study is called the
Diabetes Surveillance Dataset (DSD). It is a cleaned and consolidated version of the 2015
Behavioral Risk Factor Surveillance System (BRFSS) dataset. The DSD was curated by Alex
Teboul and can be found on Kaggle [30,35]. The diabetes surveillance dataset consisted of
21 features, which are listed in Table 1. The diabetes readmission dataset (DRD) is an edited
version of the readmission dataset from the UC Irvine Machine learning repository, which
includes ten years (1999–2008) of clinical care at 130 U.S. hospitals and integrated delivery
networks [36,37]. The diabetes readmission dataset consists of 40 features (some were re-
moved due to their complexity or missing values). Following the initial data preprocessing,
the dataset was subjected to feature selection to identify the most-relevant predictors of
diabetes and readmission risk. Statistical techniques, machine learning algorithms, and ex-
pert clinical input collaboratively informed this selection process. The resulting dataset
provided a robust foundation for the development and validation of machine learning
models aimed at projecting diabetes outcomes and readmission likelihood. These datasets
were used due to their accessibility and extensive use in the scientific community, providing
a beneficial starting point for analysis. All provided features were included in our analysis,
under the assumption that they had been previously determined as relevant for studies
of this nature. By doing so, we intended to present the full picture as showcased by the
datasets and subsequently allow the predictive models to determine the relevance of each
feature within the context of our study.

2.2. PyCaret Setup

During the setup of PyCaret in this study, we initialized the environment with a consis-
tent random state to maintain reproducibility. We then proceeded with the data preparation,
detailing the handling of missing values and the encoding of categorical variables, ensuring
clarity on the strategies employed. Feature scaling methods were applied to normalize
numerical features, essential for certain algorithms. Outliers were managed through Py-
Caret’s built-in capabilities or through additional steps, as necessary. In selecting and
training the models from PyCaret’s extensive library, we clarified our criteria, including any
baseline models used for comparison. For hyperparameter tuning, we specified the search
algorithm utilized and the hyperparameters’ range. Model evaluation and validation were
detailed by describing the use of performance metrics and validation techniques such as
cross-validation. Finally, the criteria for the final model selection were carefully outlined,
balancing performance metrics with considerations of complexity and interpretability,
to ensure a comprehensive and transparent portrayal of our research methodology.
Appl. Sci. 2024, 14, 2132 4 of 23

Table 1. Features (21) included in diabetes surveillance dataset.

Feature Type Description

Age Numeric Age of patient
Sex Nominal Sex of patient (male of female)
High blood pressure Nominal Presence or absence of high blood pressure 1
High cholesterol Nominal Presence or absence of high cholesterol 1
BMI Numeric Body mass index of patient (continuous)
Stroke Nominal Presence or absence of at least one episode of stroke 1
Heart disease Nominal History of CAD or myocardial infarction 1
Education Nominal Education level (1–6 scale)
Income Nominal Income level (1–8 scale)
Fruit Nominal Consumption of at least 1 fruit item per day 1
Vegetables Nominal Consumption of at least 1 vegetable item per day 1
Heavy alcohol Nominal Adult men ≥ 14 drinks per week and adult women ≥ 7
drinks per week 1
Smoke Nominal Smoked at least 100 cigarettes in your entire life 1
Cholesterol check Nominal Checked cholesterol in the past five years 1
Healthcare Nominal Access to any kind of healthcare 1
Cost of access Nominal Any time in the past 12 months when you needed to see a
doctor, but could not because of cost 1
Physical activity Nominal Any physical activity in the last 30 days (excluding job) 1
General health Nominal General health of patient (1–5 scale)
Mental health Numeric How many days during the past 30 days was your mental
health not good (1–30 days)
Physical health Numeric How many days during the past 30 days was your physical
health not good (1–30 days)
Difficulty walking Nominal Serious difficulty walking 1
1 Yes or no.

The PyCaret setup for both the DSD and DRD followed the same process and included
similar transformation steps. To begin, both datasets were converted to comma-separated
values (CSV) format, with each file containing a target variable. In the case of the DSD,
the target variable represented the presence or absence of diabetes. For the DRD, the target
variable had three options: no readmission, readmission within 30 days, and readmission
after 30 days. Next, the datasets were split into training and testing categories. The DSD
contained a total of 70,692 individual patient records. This dataset was divided into a
training set consisting of 49,484 patient records and a testing set consisting of 21,208 patient
records. Similarly, the DRD included a total of 101,766 patient records. This dataset was split
into a testing set of 71,236 patient records and a training set of 30,530 patient records. For the
preprocessing step applied to both the DSD and DRD, a simple imputer was used to handle
missing values in the dataset. Beyond handling missing values with a simple imputer, we
also employed scaling techniques to normalize the range of continuous input variables,
which is crucial for models that are sensitive to the varying scales of features. Additionally,
categorical variables were encoded into numerical values to ensure compatibility with
our machine learning algorithms, as they require numeric input. Outliers were treated
with careful consideration: depending on their nature and the potential impact on our
model’s ability to learn from the data, certain outliers were either adjusted or removed
to prevent them from skewing the results. Collectively, these preprocessing steps were
applied to each feature to optimize the performance of our predictive models and ensure
that any conclusions drawn from our analysis were based on data that were as accurate
and representative as possible.
Additionally, in this predictive modeling, preprocessing steps with the specific require-
ments and assumptions of the algorithms used were aligned. Missing values were handled
with imputation techniques since most algorithms necessitate complete datasets to avoid
biases and errors during training. Categorical variables were encoded to numerical values
to ensure compatibility across all models, with the encoding method chosen based on the
Appl. Sci. 2024, 14, 2132 5 of 23

algorithm’s needs; tree-based models may accommodate ordinal encoding, while others,
like logistic regression, often require one-hot encoding. Feature scaling was applied to
ensure that distance-based algorithms did not undervalue features with smaller numeric
ranges. Outlier management was crucial for preventing models sensitive to extreme values,
such as regression models, from being skewed by rare occurrences. We also employed
feature selection where necessary to mitigate the risk of overfitting and improve model gen-
eralizability. These preprocessing measures were essential in enhancing the performance
and reliability of these machine learning models, allowing us to extract the most value from
our data and ensuring robust, generalizable results.
Hence, a combination of validation techniques to ensure the reliability and accuracy of
our predictive models was employed. First, splitting the datasets into distinct training and
testing sets, providing a clear separation for unbiased model evaluation, was performed.
Furthermore, k-fold cross-validation, a method that partitions the training data into several
subsets, training the model on all but one and validating on the remaining subset, itera-
tively cycling through each subset, was utilized. This method provides a measure of the
model’s performance and its ability to generalize to unseen data. Additionally, the models’
performance using several metrics such as accuracy, precision, recall, F1-score, and the area
under the receiver operating characteristic curve, was measured. Each of these metrics
offers unique insights, together offering a comprehensive view of the model’s predictive
power. This validation framework is designed to enhance the credibility of predictive
models in this study and support their potential applicability in clinical settings.

3. Results
In this section, a comprehensive evaluation of machine learning models developed
using the PyCaret library to predict diabetes and its subsequent readmissions is presented.
Our analysis comprises two distinct subsections: the first focuses on the deployment and
performance of various algorithms in accurately predicting the incidence of diabetes mel-
litus, while the second delves into the predictive modeling of diabetes readmission rates.
Each subsection systematically illustrates the effectiveness of the employed models, dis-
cussing metrics such as the accuracy, precision, recall, and area under the receiver operating
characteristic curve. By dissecting the models’ performances in these two pivotal areas of
diabetes care, this section aims to shed light on the potential and limitations of automated
machine learning in effectively supporting disease management and healthcare delivery.
Tables 2 and 3 are structured comparisons that outline the performance of various
machine learning classifiers on diabetes prediction and readmission prediction, respec-
tively. Each classifier is evaluated according to several important metrics that gauge their
performance as follows. Accuracy: This is the ratio of the number of correct predictions to
the total number of input samples. It measures the overall correctness of the model. Higher
accuracy means the model is more often correct across all predictions. Area under the curve
(AUC): This metric is associated with the receiver operating characteristic curve and reflects
a model’s ability to distinguish between the classes (positive as diabetic and negative as
non-diabetic). A higher AUC value indicates better discrimination capabilities. Recall
(or sensitivity): Recall indicates how many of the actual positive cases (actual diabetics)
the model correctly identified. It is crucial for conditions where missing a positive case
can have serious consequences. Precision: This measures the fraction of correct positive
predictions out of all positive predictions made. In the context of diabetes, it reflects how
many of the patients the model labeled as diabetic were actually diabetic. F1-score: The
F1-score is the harmonic mean of the precision and recall, providing a balance between the
two. It is especially useful when the distribution of classes is not even, which is often the
case in medical diagnoses. Cohen’s Kappa: This measures the agreement of the predictive
model with the actual data, correcting for chance agreement. It gives a more-robust idea
of the model’s performance, especially with imbalanced datasets. Matthews correlation
coefficient (MCC): This is a measure of the quality of binary classifications. It takes into
account true positives, true negatives, false positives, and false negatives. The MCC is
Appl. Sci. 2024, 14, 2132 6 of 23

considered a balanced measure, which can be used even if the classes are of very different
sizes. Cells highlighted in yellow within Tables 2 and 3 emphasize standout performance
metrics or notable findings in the machine learning model comparison. These highlights
are used to draw the attention to results that significantly impact the study’s conclusions,
such as the highest accuracy rates, the best precision values, or any unexpected patterns in
the data that may require further investigation or discussion.

Table 2. Diabetes prediction: comparison of accuracy, area under the curve (AUC), recall, precision,
F1-score, Kappa, and Matthews correlation coefficient (MCC) for different machine learning classifier
models. Cells highlighted in yellow emphasize standout performance metrics (i.e., maximum values).

Model Accuracy AUC Recall Precision F1 Kappa MCC TT * (s)

lightgbm Light Gradient Boosting Machine 0.7536 0.8302 0.8001 0.7320 0.7645 0.5072 0.5094 2.8350
gbc Gradient Boosting Classifier 0.7526 0.8309 0.7946 0.7330 0.7626 0.5053 0.5071 4.5430
ada Ada Boost Classifier 0.7504 0.8276 0.7736 0.7392 0.7561 0.5008 0.5014 1.3560
lr Logistic Regression 0.7484 0.8247 0.7675 0.7392 0.7531 0.4968 0.4972 2.2350
ridge Ridge Classifier 0.7469 0.0000 0.7752 0.7337 0.7538 0.4928 0.4938 0.1040
lda Linear Discriminant Analysis 0.7464 0.8220 0.7775 0.7321 0.7541 0.4929 0.4939 0.2320
xgboost Extreme Gradient Boosting 0.7456 0.8205 0.7892 0.7259 0.7562 0.4911 0.4930 1.0950
qda Quadratic Discriminant Analysis 0.7294 0.7830 0.7823 0.7075 0.7430 0.4589 0.4615 0.1110
rf Random Forest Classifier 0.7254 0.7916 0.7596 0.7110 0.7345 0.4508 0.4519 5.7310
nb Naive Bayes 0.7221 0.7853 0.7143 0.7257 0.7199 0.4443 0.4443 0.0850
svm SVM-Linear Kernel 0.7216 0.0000 0.7648 0.7203 0.7257 0.4433 0.4669 1.9510
et Extra Trees Classifier 0.7093 0.7605 0.7311 0.7005 0.7155 0.4185 0.4190 6.4740
knn K Neighbors Classifier 0.7090 0.7635 0.7298 0.7008 0.7150 0.4181 0.4185 3.0830
dt Decision Tree Classifier 0.6559 0.6586 0.6368 0.6621 0.6492 0.3118 0.3120 0.2190
dummy Dummy Classifier 0.5000 0.5000 0.2000 0.1000 0.1333 0.0000 0.0000 0.0690
* TT: Training Time.

Table 3. Diabetes readmission prediction: comparison of accuracy, area under the curve (AUC),
recall, precision, F1-score, Kappa, and Matthews correlation coefficient (MCC) for different machine
learning classifier models. Cells highlighted in yellow emphasize standout performance metrics (i.e.,
maximum values).

Model Accuracy AUC Recall Precision F1 Kappa MCC TT * (s)

lightgbm Light Gradient Boosting Machine 0.5880 0.6788 0.5580 0.5586 0.5399 0.1956 0.2116 19.758
xgboost Extreme Gradient Boosting 0.5855 0.6739 0.5855 0.5540 0.5416 0.1963 0.2099 8.4470
gbc Gradient Boosting Classifier 0.5854 0.6748 0.5854 0.5574 0.5330 0.1845 0.2032 51.488
ada Ada Boost Classifier 0.5808 0.6608 0.5808 0.5452 0.5260 0.1726 0.1920 7.5600
lr Logistic Regression 0.5711 0.6426 0.5711 0.5316 0.4998 0.1333 0.1614 34.883
rf Random Forest Classifier 0.5696 0.6463 0.5696 0.5304 0.5253 0.1661 0.1769 20.993
ridge Ridge Classifier 0.5682 0.0000 0.5682 0.4934 0.4854 0.1153 0.1497 3.5300
lda Linear Discriminant Analysis 0.5680 0.6387 0.5680 0.5316 0.4933 0.1244 0.1575 5.1920
et Extra Trees Classifier 0.5523 0.6234 0.5523 0.5104 0.5131 0.1395 0.1473 25.532
dummy Dummy Classifier 0.5391 0.5000 0.5391 0.2907 0.3777 0.0000 0.0000 3.4860
svm SVM-Linear Kernel 0.5144 0.0000 0.5144 0.4911 0.4290 0.0786 0.1121 24.822
knn K Neighbors Classifier 0.4983 0.5761 0.4983 0.4793 0.4873 0.0972 0.0967 12.369
dt Decision Tree Classifier 0.4652 0.5423 0.4652 0.4703 0.4676 0.0777 0.0777 4.5380
qda Quadratic Discriminant Analysis 0.1391 0.5109 0.1391 0.4448 0.0719 0.0016 0.0046 4.1150
nb Naive Bayes 0.1259 0.5740 0.1259 0.4944 0.0561 0.0035 0.0159 3.6190
* TT: training time.

In clinical decision-making, the significance of evaluation metrics such as accuracy,

precision, recall, and AUC extends beyond their numerical definitions to their impact on
patient care, resource allocation, and overall health outcomes. Accuracy is the proportion
of true results, both true positives and true negatives, in the total population. It is a
straightforward measure and gives an overview of the model’s effectiveness. However,
in a clinical setting, high accuracy is not always indicative of practical utility, particularly
in cases where disease prevalence is low. A model could inaccurately predict true disease
cases, yet still have high accuracy due to a large number of true negatives—healthy
patients correctly identified as not having the disease. Precision, which assesses the
proportion of correctly predicted positive observations to the total predicted positives, is
Appl. Sci. 2024, 14, 2132 7 of 23

critically important when the consequences of false positives are significant. For instance,
in cancer diagnostics, a false positive might lead to unnecessary stress, invasive testing,
or treatment, which carries its own risks and costs. Recall, or sensitivity, measures
the ability of a model to find all the relevant cases within a dataset. In medicine, high
recall is crucial in conditions where failing to identify and treat a disease could lead
to severe health consequences or even death. The cost of false negatives is often much
higher than that of false positives; missing out on an early diagnosis of a disease such
as sepsis or heart attack can be life-threatening. The AUC is an aggregate measure of
performance across all possible classification thresholds. It reflects the model’s ability
to discriminate between patients with and without the condition at various levels of
specificity and sensitivity. The AUC is particularly useful in a clinical context because it
is not affected by the prevalence of the positive class, hence providing a more-balanced
evaluation of the model performance. In clinical practice, these metrics must be balanced
and interpreted in light of the disease prevalence, the potential consequences of false
positives and false negatives, and the specific clinical application. A model aimed at
diagnosing a condition with serious consequences for missed cases might prioritize
recall over precision, whereas one used to select candidates for an invasive procedure
might prioritize precision to minimize unnecessary interventions.

3.1. Diabetes Prediction

In Table 2, LightGBM performs best on most metrics, but has a slightly lower AUC
and precision compared to the Gradient Boosting Classifier and ADA/LR, respectively.
The AUC measures the model’s ability to differentiate between classes (e.g., diabetic
and non-diabetic patients). An AUC of 0.8302 for LightGBM and 0.8309 for the GBC
suggests that both models have a strong ability to discriminate between positive and
negative cases, with the GBC being marginally better. In practice, this difference might
not be clinically significant, but it indicates that the GBC could potentially provide
slightly more-accurate risk stratification. Precision assesses how many of the patients
predicted to have the condition (positive cases) actually do. LightGBM’s precision score
of 0.7320 compared to the ADA and LR’s precision score of 0.7392 suggests that, while
both models are quite good at ensuring that positive predictions are likely true, the ADA
and LR are slightly better. This means using the ADA or LR might result in fewer false
positives—patients incorrectly identified as having the condition. While LightGBM’s
lower scores in the AUC (0.8302) and precision (0.7320) compared to the GBC and
ADA/LR, respectively, mean that there are some trade-offs to consider, it is critical to
understand that machine learning predictions are not diagnostic tools themselves, but
rather, risk stratification aids that would be used in conjunction with further diagnostic
testing. Hence, LightGBM’s high performance across most metrics indicates that it is a
robust model for predicting diabetes.
Clinically, the decision on which model to use could factor in the degree of variance in
the metrics. If false positives are particularly undesirable (e.g., could lead to unnecessary
anxiety, more-invasive tests, or treatments), a physician might lean towards the ADA and
LR due to the higher precision. If the overall ability to distinguish between patients is
paramount and the differences in the precision are acceptable, LightGBM could still be
preferred given its general performance on most metrics. When these differences in scores
are very narrow (as they appear to be between these models), it becomes important to
consider other factors such as the models’ interpretability, operational costs, and how well
they integrate into clinical workflows. Clinicians might also evaluate the models based on
specific patient cohorts or different clinical settings, given that the performance could vary
across different scenarios.
The confusion matrix for the Light Gradient Boosting Machine classifier in a diabetic
prediction scenario is shown in Figure 1. The confusion matrix provides a snapshot of how
well the Light Gradient Boosting Machine classifier is performing in accurately diagnosing
diabetes as follows. The true positive rate (72%) reflects the model’s ability to identify most
Appl. Sci. 2024, 14, 2132 8 of 23

of the individuals with diabetes. The true negative rate (79%) shows that the model is also
quite capable of recognizing individuals who do not have the condition. The false positive
rate (21%) points to the rate at which the model incorrectly flags healthy individuals as
diabetic, which could lead to unnecessary worry and medical procedures for these patients.
The false negative rate (28%) is a critical metric in clinical settings, as it represents the
missed cases of diabetes that were not diagnosed by the model.

0 72% 28%
True Class

1 21% 79%

0 1
Predicted Class
Figure 1. Light Gradient Boosting Machine classifier confusion matrix for diabetes prediction.

In a clinical setting, interpreting this confusion matrix gives us insights into how well
the Light Gradient Boosting Machine classifier performs in diagnosing/predicting diabetes.
The model correctly predicted 72% of the diabetic patients. This implies that, out of all
patients who were actually diabetic, the classifier could identify 72% of them. In a clinical
setting, this is a favorable outcome since it indicates successful detection of the majority
of individuals with diabetes. The model correctly predicted 79% of non-diabetic patients.
This means 79% of the patients who did not have diabetes were correctly identified by
the model. This is important for avoiding unnecessary treatments or further medical
investigation for those who are healthy. The classifier incorrectly predicted 21% of the non-
diabetic patients as diabetic. Clinically, this could lead to unnecessary stress for the patients,
added healthcare costs, and potentially unnecessary treatment, which can have side effects.
The classifier failed to identify 28% of the actual diabetic patients, indicating that these
individuals were incorrectly classified as non-diabetic. This is particularly concerning in a
clinical environment as it means that nearly a third of diabetic patients might not receive
the necessary medical attention promptly. This could result in the progression of the disease
and associated complications since early intervention and management of diabetes are
crucial for optimal health outcomes.
In Figure 2, the ROC curves for a Light Gradient Boosting Machine classifier used
for diabetes prediction are shown, i.e., curves for each class and additional curves for the
macro-average and micro-average AUC. In the context of diabetes prediction, since there
are only two classes (diabetic and non-diabetic), Classes 0 and 1 indicate ‘non-diabetic’ and
‘diabetic’ patients, respectively. In this case, since both Class 0 and Class 1 have AUCs of
0.83, the model is equally good at identifying diabetic, as well as non-diabetic patients.
The macro-average and micro-average ROC values having AUCs of 0.83 suggest that,
overall, the model has balanced performance across both classes without any bias towards
one class due to imbalance in the dataset. In clinical settings, such a model indicates a
reliable level of discrimination for diagnosing diabetes and could, thus, be used to support
clinical decision-making.
The feature importance plot, shown in Figure 3, visualizes the significance of different
features used by the model to predict diabetes. A feature importance plot for diabetes
prediction visually represents the relative significance of various predictors in a machine
learning model’s decision-making process. In this plot, GenHlth holds the highest impor-
tance in the model, suggesting that overall self-reported health is the strongest predictor of
Appl. Sci. 2024, 14, 2132 9 of 23

diabetes among the features evaluated. Clinically, this could signify that how individuals
perceive their health is a good indicator of underlying conditions that may include or
lead to diabetes. However, general health as a feature is quite vague because it encap-
sulates various aspects of an individual’s overall well-being. This may include factors
like diet quality, physical activity levels, mental health status, and the proper manage-
ment of any existing chronic conditions. The vagueness arises from the complexity and
breadth of what constitutes “general health.” Each component (diet, exercise, etc.) alone
can have a multitude of dimensions and can be affected by subjective self-reporting and
measurement inconsistencies. Because general health is a broad concept, when it is used
as a significant feature in predictive modeling, it may provide a strong signal due to its
comprehensive nature.
1.0

0.8
True Positive Rate

0.6

0.4
ROC of Class 0, AUC = 0.83
ROC of Class 1, AUC = 0.83
0.2
Macro-Average ROC, AUC = 0.83
Micro-Average ROC, AUC = 0.83
0.0
0.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate

Figure 2. ROC curves for a Light Gradient Boosting Machine classifier used for diabetes prediction
(diagonal line in the graph indicates that lines above it are significant or noteworthy).

With about half the importance value of GenHlth, high blood pressure is also a signifi-
cant predictor, but less so than general health status. This reflects the known association
between hypertension and the risk of developing diabetes [38]. BMI, age, and HighChol:
each of these features has an importance value around one quarter that of GenHlth, indicat-
ing moderate predictive power. These factors are well-established risk factors for diabetes:
a higher BMI can indicate obesity [39]; age is a known risk factor as risk increases with
age [40]; high cholesterol levels are associated with increased cardiovascular risks and can
be indicative of metabolic issues related to diabetes [41]. Income, HeartDiseaseOrAttack,
DiffWalk, HighAlcoholConsumption, CholCheck: These features have low, but non-zero
importance values, which means the model found them to have some predictive value,
but they contribute far less to the prediction than the top factors.

GenHlth
HighBP
BMI
Age
Features

HighChol
Income
HeartDiseaseOrAttack
DiffWalk
HighAlcoholConsumption
CholCheck
0.0 0.08 0.24 0.16 0.32 0.40
Variable Importance

Figure 3. Feature importance plot for diabetes prediction.

In clinical application, a plot like this helps practitioners and researchers concentrate
their efforts on the most-telling risk factors when developing preventative measures and
Appl. Sci. 2024, 14, 2132 10 of 23

management strategies for diabetes. For example, interventions may focus on improving
general health, since this is the most-influential feature; despite its vagueness, healthcare
providers might emphasize comprehensive health improvement strategies that encompass
diet, exercise, mental health, and chronic disease management. Managing high blood
pressure: as the second-most-important factor, effective hypertension management could
be a key focus for reducing diabetes risk. This could involve medication, dietary changes,
and lifestyle modifications. Addressing obesity: given the BMI’s role as a moderate
predictor, weight management programs could be an important component of diabetes
prevention efforts. Monitoring and support for aging populations: as age is a predictor,
targeted screening and intervention programs could be developed for older individuals.
Controlling cholesterol levels: since high cholesterol is a moderate predictor, cholesterol
management through diet, exercise, and possibly medication could also be a priority. While
the features with lower importance values contribute less to the model, this does not mean
they should be ignored in clinical practice. Income, for instance, can influence health
outcomes through complex social determinants of health. Heart disease, physical difficulty
in walking, alcohol consumption, and cholesterol checkups are all parts of a patient’s
comprehensive health profile and, while less predictive in this model, still play roles in
overall health and diabetes risk.
Figure 4 shows the learning curves for the Light Gradient Boosting Machine classifier
used to predict the diabetes. It shows the testing score starting above the cross-validation
score and, then, declining to approach the cross-validation score as the number of training
instances increases. This pattern suggests that, as the model is exposed to more varied data
during training, it begins to generalize better. The initial high testing score might be a result
of the model being too tailored to a small dataset and not capturing the true underlying
patterns. When both scores converge, this indicates that adding more data is helping the
model to stabilize. The testing score decreases to meet the cross-validation score, and this
convergence signifies that the model is now generalizing well and its performance on the
unseen test data is becoming consistent with the validation data used during training.
A model that generalizes well is crucial in clinical settings because it indicates that the
model can perform reliably when diagnosing new patients not seen during the training
phase. The learning curve can also suggest whether more data could help to improve
the model or whether the model has plateaued and additional data might not increase
performance. Since the curve have not plateaued in this case, this indicates that the model
has not reached its peak performance and additional training with more data would help
to improve it.

0.761 Training Score

Cross-Validation Score
0.758

0.755
Score

0.752

0.749

0.746

15,000 24,000 33,000 42,000

Training Instances

Figure 4. Learning curves for Light Gradient Boosting Machine classifier to predict diabetes.

The validation curves in Figure 5 are a graphical representation showing the

model performance (score) on both the training and validation datasets as a function
of ‘max_depth’ for the Light Gradient Boosting Machine classifier. The following in
the validation curves for max_depth in a Light Gradient Boosting Machine classifier
used to predict diabetes can be observed. Initially, when the max_depth is at its lowest,
Appl. Sci. 2024, 14, 2132 11 of 23

both the training and cross-validation scores are similar and low. This indicates that the
model is underfitting the data; it is not complex enough to capture the underlying rela-
tionships within the data. However, as the max_depth increases, the model is allowed
to create more-complex decisions, and it starts to fit the training data better, resulting
in an increased training score. However, the cross-validation score begins to decrease
after a certain point, suggesting the model is starting to overfit the training data. This
divergence is a classic sign of overfitting. The classifier, with increasing max_depth, is
becoming more specialized in dealing with the intricacies of the training data, including
noise, which does not generalize well to unseen data. In practice, for predicting diabetes,
this pattern means that an ideal max_depth should balance the training and validation
scores. It should be high enough to adequately learn from the data, but not so high as
to overfit. Tuning would involve finding a max_depth where the validation score is
maximized before it starts to decrease. The ideal max_depth would provide a model that
is adequately complex to predict diabetes accurately, but not too complex that it does not
generalize well to new patient data. A model that overfits may perform exceptionally
on historical data, but fail in a real-world clinical setting. This could lead to incorrect
predictions, which, in turn, could potentially cause harm if the predictions are used to
guide treatment or diagnose patients. Hence, when using a Light Gradient Boosting
Machine classifier for diabetes prediction, one must be careful not to choose a max_depth
that is too high to avoid overfitting.

0.88
Training Score
0.85
Cross-Validation Score
Score

0.82

0.79

0.76

2 4 6 8 10
max_depth

Figure 5. Validation curves for Light Gradient Boosting Machine classifier used to predict diabetes.

3.2. Diabetes Readmission Prediction

In Table 3, the Light Gradient Boosting Machine classifier scores highest in all
metrics except for F1 and Kappa, where the Extreme Gradient Boosting performs best.
In the context of diagnosing diabetes, high accuracy indicates that LightGBM makes the
most-correct predictions across all cases evaluated. This is generally desirable, but high
accuracy alone can be misleading if the dataset is imbalanced (i.e., there are significantly
more non-diabetic than diabetic patients). The high AUC value for LightGBM suggests
it is good at distinguishing between diabetic and non-diabetic patients. The fact that
XGBoost has the highest F1-score suggests that it has a better balance between precision
and recall compared to LightGBM, which may be particularly relevant if one is more
important than the other—for example, if it is crucial to minimize false negatives due
to the severe consequences of undiagnosed diabetes. The higher Kappa for XGBoost
indicates a stronger agreement between the predicted and actual values, considering
the possibility of random chance leading to agreement. Hence, even though LightGBM
shows the highest performance in most metrics, XGBoost stands out in terms of F1 -score
and Kappa. These suggest that XGBoost, while potentially slightly less accurate overall,
might offer a more-reliable balance of precision and recall, making fewer mistakes for
both false positives and false negatives, which in medical applications can be just as
important as overall accuracy.
Appl. Sci. 2024, 14, 2132 12 of 23

In practice, the choice between the models may come down to the specific clinical
requirements and the costs associated with different types of diagnostic errors. In the
end, the choice between these models would likely involve a trade-off depending on
whether sensitivity or precision is more valued for the particular application of diagnosing
diabetes. For example, in a screening context where false negatives could be more harmful,
a model with a higher recall might be more appropriate. It is also possible that different
healthcare settings might deploy different models accordingly; for instance, one might use
LightGBM in general screening due to its high accuracy, and another might use XGBoost in
high-risk populations where the costs of false negatives are particularly high. However,
the F1 and Kappa values for LightGBM are very close to those for XGBoost, which has the
highest values in these metrics. This suggests that LightGBM provides a near-equivalent
balance of precision and recall and almost as much reliability and agreement as XGBoost.
With LightGBM scoring slightly lower on the F1-score and Kappa, but still in close range to
XGBoost, we favor LightGBM because it outperforms XGBoost in other important metrics
like accuracy and AUC.
Figure 6 presents a confusion matrix, i.e., a tabular representation illustrating the
performance of the Light Gradient Boosting Machine classifier in predicting diabetes
readmission. The three conditions represented are as follows. Condition 0: no readmission,
Condition 1: readmission before 30 days, and condition 2: readmission more than 30 days
after discharge. The confusion matrix elements, as denoted by x_ij, where i is the true
class and j is the predicted class, quantify the percentage of predictions that fall into
each category as follows. x11 = 3%: percentage of patients correctly predicted as ‘no
readmission’ (true negatives for condition 0). x21 = 1%: percentage of patients incorrectly
predicted as ‘no readmission’ when they were actually readmitted before 30 days. x31 = 0%:
percentage of patients incorrectly predicted as ‘no readmission’ when they were actually
readmitted after more than 30 days. x12 = 38%: Percentage of patients incorrectly predicted
as ‘readmitted before 30 days’ when there was no readmission. x22 = 37%: percentage of
patients correctly predicted as ‘readmitted before 30 days’ (true positives for condition 1).
x32 = 15%: percentage of patients incorrectly predicted as ‘readmitted before 30 days’
when the readmission actually occurred after more than 30 days. x13 = 59%: percentage
of patients incorrectly predicted as ‘readmission after more than 30 days’ with no actual
readmission. x23 = 63%: percentage of patients incorrectly predicted as ‘readmission
after more than 30 days’ when they were actually readmitted before 30 days. x33 = 84%:
percentage of patients correctly predicted as ‘readmission after more than 30 days’ (true
positives for condition 2).

0 3% 38% 59%
True Class

1 1% 37% 63%

2 0% 15% 84%

0 1 2
Predicted Class
Figure 6. Light Gradient Boosting Machine classifier confusion matrix for readmission prediction.

Clinically, in practice, this confusion matrix indicates how accurately the model
can predict patient readmission, which is critical information for healthcare providers
and institutions. The ability to predict readmissions can help in targeting interventions
to reduce preventable readmissions and improve patient outcomes. The data can be
interpreted clinically as follows. The high true positive rate for condition 2 (84%) suggests
Appl. Sci. 2024, 14, 2132 13 of 23

that the model is particularly effective at identifying patients at risk of readmission more
than 30 days after discharge, which could be valuable for long-term patient care planning
and intervention. The relatively low true positive rates for conditions 0 and 1 (3% and
37%, respectively) indicate that the model is less reliable at classifying patients who are
not likely to be readmitted, as well as those who might be readmitted within 30 days.
This might lead to some degree of unnecessary resource allocation for individuals less
at risk while potentially under-serving those at higher risk of short-term readmission.
The incorrect predictions, such as patients without readmission being classified as
readmitted before 30 days (38%) or after more than 30 days (59%), could result in
unnecessary follow-up, adding to patient anxiety and burdening the healthcare system.
Conversely, the model’s incorrect predictions in missing true cases of readmission within
30 days (i.e., x21 and x23) represent a risk of not providing necessary care for those
patients, which could lead to worse health outcomes. In practice, these results would
necessitate a balance between managing the risks of unnecessary follow-ups (resulting
from false positives) versus the risk of missing early interventions (due to false negatives).
This balance is important for the quality of patient care, clinical workflow, resource
allocation, and healthcare costs.
Figure 7 illustrates the ROC curves for a Light Gradient Boosting Machine classifier
that is used for predicting diabetes readmission. Each ROC curve corresponds to one of
the readmission categories, and the area under each curve provides a measure of how
well the model distinguishes between classes. The ROC of class 0, AUC = 0.67: This is
the ROC curve for patients with no readmission. An AUC of 0.67 suggests that the model
has a moderate ability to distinguish between patients who will not be readmitted and
those who will. ROC of class 1, AUC = 0.66: This curve pertains to patients readmitted
before 30 days. The AUC value indicates that the classifier’s ability to differentiate between
those readmitted within 30 days and all other outcomes is slightly less than moderate.
ROC of class 2, AUC = 0.69: This relates to the patients readmitted more than 30 days after
discharge. An AUC of 0.69 shows that the model is moderately good at differentiating
between late readmissions and the other outcomes. Macro-average ROC, AUC = 0.78: The
macro-average AUC is calculated by taking the average of the AUC values for each class.
An AUC of 0.78 is relatively high, indicating good overall discriminatory ability across
all classes without taking class imbalance into account. Micro-average ROC, AUC = 0.67:
The micro-average AUC aggregates the contributions of all classes to compute the average
AUC. A value of 0.67 suggests that the model has a moderate overall performance at
distinguishing between classes when considering class imbalance.
1.0

0.8
True Positive Rate

0.6

0.4
ROC of class 0, AUC = 0.67
ROC of class 1, AUC = 0.66
0.2 ROC of class 2, AUC = 0.69
macro-average ROC, AUC = 0.78
micro-average ROC, AUC = 0.67
0.0
0.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate
Figure 7. ROC curves for a light gradient boosting machine classifier used for diabetes readmis-
sion prediction.

Clinically, the AUC values provide insight into the model’s predictive power. An AUC
closer to 1 indicates excellent model performance, while an AUC closer to 0.5 suggests that
the model is no better than random chance. In this case, the LightGBM classifier shows
moderate predictive performance for individual classes, and the macro-average indicates a
Appl. Sci. 2024, 14, 2132 14 of 23

better overall performance. It is important for clinicians and healthcare administrators to

consider these values when deciding how to incorporate the model into practice. When
using this model, it may need to be combined with other clinical assessments and risk
stratification tools to make more-informed decisions about patient care following discharge.
The feature importance plot for diabetes readmission prediction is shown in Figure 8.
The plot shows that AvgGlucoseLevel holds the greatest importance. This implies that
the average blood glucose level of a patient is the strongest predictor for whether a
patient with diabetes might be readmitted to the hospital. It is clinically meaningful
as glucose levels need to be managed effectively in diabetics to prevent complications.
BMI has the second-highest importance value. This reflects that a patient’s body weight
relative to his/her height is a significant factor in readmission risk, likely due to obesity
being a risk factor for diabetes complications. Lower values are attributed to features
like WorkType, EverMarried, Hypertension, SmokingStatus, HeartDisease, and Res-
idenceType. Although these have an association with the risk of readmission, their
influence is stronger than factors like sex and age, but less than AvgGlucoseLevel and
BMI. These could represent lifestyle and social determinants of health that impact patient
outcomes to a lesser degree. Finally, sex and age are given zero values, suggesting that,
in this particular model, they do not contribute to predicting readmission for patients
with diabetes. This does not mean these factors are not important in other contexts,
but in this specific model’s configuration, their impact is non-existent or negligible. This
hierarchy in feature importance can help healthcare providers focus their interventions
and follow-up strategies on what matters most for preventing readmissions in diabetic
patients. They may focus more on controlling blood glucose levels and managing weight,
while still being aware of other factors that could influence a patient’s risk of returning
to the hospital.

AvgClucoseLevel
BMI
WorkType
EverMarried
Features

Hypertension
SmokingStatus
HeartDisease
ResidenceType
Sex
Age
0.0 0.07 0.14 0.21 0.28 0.35 0.42
Variable Importance
Figure 8. Feature importance plot for diabetes readmission prediction.

In Figure 9, the following characteristics in the learning curves for a Light Gradient
Boosting Machine classifier used for diabetes readmission prediction can be observed. Since
the training curve is constant at 1, it suggests that the LightGBM model is perfectly fitting
the training data throughout the training process, from the smallest to the largest number
of training instances. A constant score of 1 indicates that the model has 100% accuracy on
the training dataset, which often points to overfitting. The increasing of the cross-validation
score curve signifies that, as more data are provided for training, the model’s performance
on unseen data (the validation set) is improving. Initially, when the number of training
instances is small, the model may not have learned the underlying patterns adequately,
which is reflected in a lower cross-validation score. As training progresses and the model
is trained on more data, it begins to generalize better, capturing the essential patterns to
make more-accurate predictions on unseen data. The cross-validation score approaching
1 towards the end of the training process means that the model’s predictive performance
on unseen data is nearly perfect. This can be an indicator that the model has learned the
Appl. Sci. 2024, 14, 2132 15 of 23

underlying trends in the data quite well and is able to generalize this learning to new
data effectively.

1.00

0.95

Score 0.90

0.85
Training Score
0.80
Cross-Validation Score
0.75

8,000 14,000 20,000 26,000

Training Instances
Figure 9. Learning curves for light gradient boosting machine classifier used for diabetes readmis-
sion prediction.

In the clinical context, learning curves with such characteristics have both positive
and negative implications. A perfect training score is often a red flag for overfitting.
In a clinical setting, an overfitted model may perform exceptionally well on historical
data, but fail to predict accurately on future patients or those from different demographic
groups. Therefore, the clinical reliability of such a model might be questionable. However,
the increasing cross-validation score is a good sign that the model is improving its ability
to generalize as it sees more data. Clinically, this can translate to more-reliable predictions
for patient readmissions as the model is exposed to and learns from a broader variety of
patient cases. Therefore, before the model is used in practice, it would be necessary to
validate its performance on an independent test set that the model has never seen before.
If the model performs equally well on the independent test set, this would provide more
confidence that it will perform well in a real-world clinical environment. Additionally,
if the cross-validation score continues to improve and reaches nearly 1, it could suggest that
the model training has reached a plateau; additional training instances might not lead to
any significant improvements. Clinically, if the validation and independent testing confirm
these results, this model could be considered reliable for predicting diabetes readmission.
The validation curves for the Light Gradient Boosting Machine classifier used to
predict diabetes readmission are shown in Figure 10. When both the training score and
the cross-validation score of the Light Gradient Boosting Machine Classifier used for
predicting diabetes readmission increase together as the ‘max_depth’ hyperparameter
increases and neither reaches a plateau, this can be indicative of a few things as follows.
Underfitting at lower ‘max_depth’: Initially, both scores starting at the same lower value
typically means that, at low complexity (shallower trees), the model may not be capturing
enough of the data’s complexity to make accurate predictions. Increasing complexity
improves performance: As ‘max_depth’ increases, allowing the model to represent more-
complex functions, both the training and validation scores improve. This behavior indicates
that more-complex models are doing a better job of capturing the trends in the data
without fitting to noise and, hence, are generalizing well. No evidence of overfitting with
increased ‘max_depth’: The fact that the validation score continues to increase suggests
that the model has not yet begun to overfit to the training data. Overfitting is typically
marked by a decrease in the validation score as the complexity of the model increases
beyond a certain point. No plateau has been reached: Neither the training score nor the
validation score leveling off suggests that the model might benefit from further increases
in complexity, and there might still be room to improve the model’s performance with
increasing ‘max_depth’ up to a certain point that has not been reached within the range of
the ‘max_depth’ values considered.
Appl. Sci. 2024, 14, 2132 16 of 23

0.85
Training Score
0.80
Cross Validation Score

Score
0.75

0.70

0.65

2 4 6 8 10
max_depth
Figure 10. Validation curves for Light Gradient Boosting Machine classifier used to predict dia-
betes readmission.

4. Discussion
Diabetes is a multifaceted disease, influenced by genetics, lifestyles, and socioeconomic
factors, thus demanding personalized treatment approaches. Machine learning models
stand at the forefront of enabling such individualized care by leveraging vast amounts
of health data to uncover complex, nonlinear relationships that may go undetected by
traditional methods [42,43]. The importance of these models lies in their ability to digest
diverse datasets, including genetic profiles, dietary habits, exercise routines, and even
environmental and social influences, to predict diabetes-related outcomes with precision.
By accommodating the unique aspects of each patient’s condition, machine learning facili-
tates a more-nuanced understanding of the disease, paving the way for tailored treatment
plans. Such customization is increasingly necessary to manage diabetes effectively, given its
dynamic nature and varying impact across populations, reinforcing the essential role that
machine learning plays in advancing diabetes care. In this study, we employed the PyCaret
machine learning library to predict diabetes outcomes and readmission rates, a task of high
clinical relevance given the burden of diabetes on patients and healthcare systems. Our
comparative assessment of advanced algorithms underscores the intricate balance between
sensitivity and specificity in medical diagnostics.
The results presented confirmed the capacity of the PyCaret library to facilitate the
development of robust predictive models. The Extra Trees classifier, among others, emerged
as a prominent algorithm owing to its exceptional performance metrics such as the Kappa
statistic and Matthews correlation coefficient. These metrics highlight the classifier’s
proficiency in identifying true positive and negative cases, a characteristic that is essential
in clinical decision-making to minimize the risk of overlooking diabetes diagnoses or
inappropriately predicting readmissions. The application of transfer learning and the
utilization of pretrained models present a promising approach when working with limited
datasets. These methods leverage the knowledge gained from large, previously trained
models, adapting it to smaller, domain-specific datasets with potentially fewer instances.
By fine-tuning the last layers of a pretrained network, one can enhance performance on
a new task without the need for the extensive data that are often required to train deep
learning models from scratch. This technique can capitalize on the generalizability of
features learned from the larger dataset, allowing for the efficient utilization of the available
data and potentially yielding improved performance on tasks with limited samples. It
opens the door for incremental learning, where the model can be continuously updated as
new data become available, making it highly suitable for dynamic fields such as healthcare,
where obtaining large, annotated datasets can be challenging.
Machine learning models that have proven effective in our study are likely to have far-
reaching implications. For healthcare professionals, the application of these models could
translate to a transformative approach in the prognosis of diabetes and its complications,
with the potential to significantly reduce readmission rates through precise and timely
diagnosis. This could lead to more-individualized patient care, optimizing the use of
Appl. Sci. 2024, 14, 2132 17 of 23

healthcare resources and potentially improving the quality of life for individuals with
diabetes. Furthermore, our analysis shed light on the risk factors predictive of diabetes
and its associated readmissions. Such insights are pivotal as they can inform targeted
interventions for high-risk populations. However, it is paramount to note that, while the
models exhibit substantial predictive power, model refinement and validation in diverse
clinical settings are imperative to bolster their applicability. Surprisingly, the study revealed
disparities in the predictive models’ ability to detect early versus late readmission. This
highlights an opportunity for further investigation into model tuning to enhance early
readmission predictions, which are crucial for immediate post-discharge care and planning.
This study utilizes open-source datasets, specifically the DSD and the DRD from the
UC Irvine Machine Learning Repository, due to their wide use and accessibility within the
research community. It is important to note, however, that the reliance on these datasets
introduces certain limitations regarding the quality and the representativeness of the data.
Such datasets may not capture the full spectrum of the diabetic population, potentially
impacting the generalizability of our model’s findings. The representativeness of the pa-
tient data from these datasets, as compared to the broader diabetes population, should
be critically considered when interpreting the results of this study. Further research us-
ing a diverse range of datasets, including those that more accurately reflect the current
demographic makeup of the diabetic community, would be beneficial for confirming the
validity and improving the applicability of our predictive modeling approach. This is a
recognized challenge within the field and one that is important for the advancement of
machine learning tools aimed at diagnosing and managing diseases like diabetes. Addi-
tionally, the discussion of overfitting reflects a recurrent challenge in machine learning.
Our learning and validation curves indicated the necessity for optimal complexity in our
models. The drive to improve sensitivity must be carefully weighed against the risk of
generating models too finely tuned to training data that they fail to generalize to new
patients. Therefore, continuous refinement of the balance between model complexity and
generalizability is essential. Hence, while the application of the PyCaret library in predict-
ing diabetes and readmission has shown promise, further exploration and validation are
required. The inescapable trade-offs between different performance metrics point to the
necessity for an interdisciplinary approach that combines machine learning expertise with
clinical knowledge. The efficacious application of a Light Gradient Boosting Machine classi-
fier for diabetes prediction and diabetes readmission prediction in our analysis underscores
this synergy. These models have exhibited a balance between capturing complex patterns
in the data and avoiding the pitfall of overfitting.
Our findings suggest that the Light Gradient Boosting Machine classifier, with its
capability to handle various types of features and data distributions, stands out for its
diagnostic accuracy for diabetes. This has broad implications for the early identification of
diabetic patients, which is crucial for the timely initiation of the treatment and prevention
of acute complications. The model’s performance was particularly strong in distinguishing
between diabetic and non-diabetic cases, a result that may be attributed to its iterative
approach, which focuses on correcting the misclassified observations in successive training
rounds. However, machine learning models with a 20–30% false positive rates indicate
that the models misclassified about one-quarter of the negative cases as positive, such as
predicting that patients will be readmitted when they will not be. This could be due to a
variety of factors, including imbalanced datasets, where one class predominates, leading
to bias; inadequate or noisy features that fail to capture the essence of the predictive
patterns; improperly calibrated decision thresholds that tip the balance towards positive
predictions; overfitting, where the model fits the training data including their noise too
closely and fails to generalize; or underfitting, where the model is too simplistic to capture
the underlying structure of the data. It might also be attributed to data quality issues, such
as inaccuracies in the data that the model was trained on. To reduce false positives in
the future, efforts should focus on ensuring high-quality data and possibly reevaluating
the features used for prediction. Consideration might also be given to employing class
Appl. Sci. 2024, 14, 2132 18 of 23

balance techniques, adjusting the decision-making threshold, and applying regularization

techniques to improve the model’s generalizability. These steps are crucial for making
the model more reliable and, thus, more useful in clinical applications where accurate
predictions can significantly impact patient outcomes.
The models discussed in the study can uncover the influence of feature interrelations
that might not be evident when considering individual feature’s impacts in isolation.
Machine learning techniques, especially those encompassing complex models like ensemble
methods or deep learning, are capable of capturing nonlinear interactions and dependencies
between features. For instance, a single variable might have minimal predictive power
on its own, but when combined with others, it can contribute significantly to the model’s
accuracy. By examining feature importance plots and analyzing the model’s decision-
making process, insights into complex interactions can be deduced. Understanding these
interrelations is critical because it can lead to more-accurate predictions and can inform
clinicians about synergistic or antagonistic effects between variables. This might lead to
the discovery of previously unknown patterns in the data, influencing decisions regarding
treatment plans or follow-up strategies. It highlights the need for careful consideration of
the model’s interpretability and the use of appropriate tools to shed light on how different
features influence model predictions.
It is essential to focus on the potential impact of these models on patient outcomes
and the healthcare system in general. These models must deliver actionable insights that
improve the accuracy and timeliness of diagnoses, guiding interventions that mitigate the
risk of complications and readmissions. Clinicians rely on the outputs of these models
to make decisions, so the models must be trustworthy and their predictions should align
with real-world clinical scenarios. The effectiveness of a predictive model is not solely
determined by its statistical accuracy; it also depends on its ability to support healthcare
providers in making decisions that lead to better care and resource utilization. A model
that accurately identifies patients at high risk for diabetes complications or readmission can
help in allocating medical resources more effectively and possibly preventing worse health
outcomes. Conversely, a model must also avoid overburdening the healthcare system with
false alerts that could lead to unnecessary tests or treatments. Ultimately, incorporating
predictive models into clinical practice involves a careful consideration of these trade-offs
to ensure that the models serve their intended purpose—to support clinicians in delivering
optimal care to diabetic patients and efficiently manage hospital resources.
For predicting diabetes readmission, the metrics achieved by the Light Gradient Boost-
ing Machine classifier indicate its utility in prognostic settings. The ability to accurately
predict readmissions can profoundly affect patient management, potentially preventing
the adverse outcomes associated with early rehospitalization. This outcome is particularly
vital as it enables healthcare providers to identify individuals most at risk and deploy in-
terventions preemptively, thus enhancing patient care and optimizing resource utilization.
The dynamic nature of diabetes progression and readmission rates demands that predictive
models are not only accurate, but also adaptable. Our learning curves for the Light Gradient
Boosting Machine classifier point towards a model that benefits from more data. Similarly,
the validation curves suggest that models need to be finely tuned to generalize well to
unseen patient data undefined. This need for adaptability raises the question of how the
models would perform when subjected to data from varied demographics or different
healthcare settings.
The clinical implications of our study are far-reaching. From guiding policy on diabetes
management programs to influencing the allocation of resources for patient education and
follow-up care, the potential applications of our research are myriad. One of the most-
critical considerations moving forward would be the ethical implementation of these
predictive models, ensuring that they serve to augment, rather than replace, the clinical
judgment of healthcare professionals. In light of the promising performance exhibited by
the predictive capabilities of the Light Gradient Boosting Machine classifier, it remains
essential to examine the broader implications of integrating these models into clinical
Appl. Sci. 2024, 14, 2132 19 of 23

workflows. The Light Gradient Boosting Machine classifier, with its high-performance
metrics, offers a compelling solution for diabetes prediction. This model, which excels in
various performance metrics, can augment the clinician’s ability to diagnose diabetes more
effectively, potentially facilitating better patient outcomes through earlier intervention and
targeted treatment plans.
The impactful performance of the Light Gradient Boosting Machine classifier in pre-
dicting diabetes readmission emphasizes its potential role in post-discharge care. Accurate
predictions of readmission can enable healthcare professionals to identify at-risk individu-
als and tailor follow-up care accordingly. This could have substantial benefits in terms of
reducing hospitalization costs, enhancing patient satisfaction, and, most importantly, im-
proving the health status of patients with diabetes. The presented learning curves indicated
that the models’ performance could be further enhanced by increasing the volume and
diversity of the training data. This points towards a growing need for extensive datasets
that reflect the multifaceted nature of diabetes across different populations and healthcare
settings. The validation curves, on the other hand, suggest that the selected hyperparame-
ters offer a satisfactory balance between learning the training data’s underlying trends and
maintaining generalization over the validation data.
It is also worth considering the practical applicability of our findings in the healthcare
industry. For example, the deployment of these models on electronic health records could
streamline the identification of patients requiring closer management, thereby personal-
izing healthcare delivery. Nonetheless, we must also be mindful of the risks associated
with over-reliance on algorithmic decisions. It is imperative to integrate these models
within a decision support framework that respects the nuances of individual patient cases
and leverages the indispensable expertise of healthcare providers. The uncertainties of
predictive models, particularly in the context of diagnosing diabetes and predicting hos-
pital readmissions, stem from various sources and can significantly impact their clinical
utility. Firstly, there is the inherent variability in patient populations. The datasets used to
train these models may not capture the full spectrum of diabetes presentations, leading to
uncertainties when the models are applied to diverse real-world populations. Changes in
population demographics, disease prevalence, and practice patterns can all influence model
performance. Secondly, data quality and completeness can introduce uncertainty. Models
are developed based on available data, which might have missing values, measurement
errors, or lack critical predictors not captured in the dataset. This limitation can impact
the precision of the predictions, leading to a higher number of false negatives or positives.
Thirdly, the choice of modeling approach and complexity of the model can lead to different
levels of uncertainty. Complex models might perform excellently on training data, but may
not generalize well to new data (overfitting), while simpler models may not capture all
the nuances of the disease (underfitting). Another source of uncertainty is the evolution
of medical knowledge and practices. As new clinical guidelines and treatments emerge,
the relevance and effectiveness of existing models may diminish if they are not updated
accordingly. Furthermore, the interpretability of the models also introduces uncertainties in
clinical practice. If the decision-making process of the model is not transparent, clinicians
might be uncertain about the reliability of the predictions, which could hinder their adop-
tion. Lastly, patient behavior and adherence to treatment can introduce unpredictability
that is difficult for models to account for. Factors such as lifestyle changes, medication
adherence, and access to healthcare services can significantly sway the outcomes that
models attempt to predict. All these aspects contribute to the overall uncertainty associated
with predictive models and need to be considered when integrating them into clinical
workflows, communicating risks to patients, and making informed decisions based on
their outputs.
Hence, the study affirms the value of PyCaret-based machine learning models in
addressing complex clinical questions, such as diabetes prediction and readmission rates.
While the results are promising, continuous collaboration between data scientists and
healthcare professionals is essential in refining these models. Future research should focus
Appl. Sci. 2024, 14, 2132 20 of 23

on external validation, exploring the utility of these models across various patient demo-
graphics and healthcare infrastructures, ensuring their reliability and helping to bridge the
gap between machine learning and patient care. The pursuit of such interdisciplinary col-
laborations holds the key to unlocking the full potential of machine learning in enhancing
healthcare outcomes for patients with diabetes.
In looking to the future of diabetes prediction models, the emphasis must be placed on
enhancing clinical applicability, accommodating the intricacies of personalized healthcare
and ensuring adaptability across varying demographics and healthcare infrastructures.
Clinically applicable models need to be user-friendly and easily integrated into healthcare
providers’ existing routines, enabling prompt and informed decisions. Furthermore, en-
suring that these predictive tools are relevant and effective for diverse populations and
can operate within different healthcare setups, from urban hospitals to resource-limited
rural clinics, is crucial for their global utility. Investing in these areas will help to refine
diabetes management and contribute to overall improvements in healthcare outcomes.
The pursuit of AI transparency is a key trajectory for future developments, and it holds
particular significance in the healthcare sector, where the decision-making process must
be clear and justifiable. The opacity of complex models, often referred to as ‘black-box’
systems due to their obscure internal workings, can hinder their credibility and uptake
by medical professionals who require a thorough understanding of how predictions are
arrived at to inform patient care. Regulatory entities are increasingly emphasizing the need
for transparency in AI tools destined for clinical applications. As certainty and acceptance
grow with more-intelligible models, the articulation of AI decision processes is expected to
become a standard requisite. This will not only assist healthcare professionals, but will also
ensure that such systems adhere to ethical standards and regulations.

5. Conclusions
Provided below is a concise summary of the study’s principal findings and viewpoints,
structured as bullet points. These encapsulate the study’s contributions to the use of ma-
chine learning in healthcare, highlighting the capability to improve diabetes management
and identifying paths for subsequent research:
• Diabetes mellitus is a metabolic condition that leads to high blood glucose levels,
increasing the risk of cardiovascular disease and requiring a multidisciplinary man-
agement approach.
• The study explores the use of machine learning, specifically the PyCaret library,
in predicting and managing diseases, with a focus on diabetes mellitus and hospital
readmission rates.
• Machine learning has been used to predict the incidence and risk factors of diabetes
and other cardiovascular diseases, using a dataset from the Kaggle repository.
• The Light Gradient Boosting Machine classifier was identified as a prominent algo-
rithm due to its high performance metrics, particularly in predicting late readmissions.
• However, the model was less reliable in classifying patients who are not likely to be
readmitted or those who might be readmitted within 30 days.
• There is a need for further research and validation across diverse clinical settings to
fully incorporate machine learning into modern diabetes care.
• The study highlights the importance of balancing different performance metrics and
considering the specific clinical requirements and costs associated with different types
of diagnostic errors.
• The study affirms the value of machine learning models in addressing complex clinical
questions and improving the accuracy and timeliness of diagnoses.
• The findings ought to be contextualized within the broader landscape of healthcare
analytics and their implications for diabetes management.
• Future research should focus on external validation and explore the utility of these
models across various patient demographics and healthcare infrastructures.
Appl. Sci. 2024, 14, 2132 21 of 23

Author Contributions: Conceptualization, R.J., F.S., A.T. and M.T.; methodology, R.J., F.S., A.T. and
M.T.; software, R.J., F.S., A.T. and M.T.; validation, R.J., F.S., A.T. and M.T.; formal analysis, R.J., F.S.,
A.T. and M.T.; investigation, R.J., F.S., A.T. and M.T.; resources, R.J., F.S., A.T. and M.T.; data curation,
R.J., F.S., A.T. and M.T.; writing—original draft preparation, R.J., F.S., A.T. and M.T.; writing—review
and editing, R.J., F.S., A.T. and M.T.; visualization, R.J., F.S., A.T. and M.T.; supervision, M.T.; project
administration, M.T. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The datasets analyzed in this study, as well as the software used for
their analysis, are available online. The corresponding references to these datasets and software are
included in the article. Any additional information or data not provided in the article can be obtained
upon reasonable request.
Conflicts of Interest: The authors declare no conflicts of interest.

Abbreviations
The following abbreviations are used in this manuscript:

CVD cardiovascular disease

BMI body mass index
CAD coronary artery disease
DSD Diabetes Surveillance Dataset
BRFSS Behavioral Risk Factor Surveillance System
DRD diabetes readmission dataset
AUC area under the curve
MCC Matthews correlation coefficient
TT training time
ROC receiver operating characteristic

References
1. Cloete, L. Diabetes mellitus: An overview of the types, symptoms, complications and management. Nurs. Stand. 2021, 37, 61–66.
[CrossRef]
2. American Diabetes Association. Diagnosis and Classification of Diabetes Mellitus. Diabetes Care 2009, 32, S62–S67. [CrossRef]
[PubMed]
3. Negroiu, C.E.; Tudorascu, I.; Moise, C.G.; Vinturis, E.; Bezna, C.M.; Danoiu, R.; Gaman, M.E.; Danoiu, S. Obesity and Myocardial
Infarction-The Place of Obesity Among Cardiovascular Risk Factors-Retrospective Study. Curr. Health Sci. J. 2023, 3, 388–396.
[CrossRef]
4. Desai, R.; Katukuri, N.; Goguri, S.R.; Kothawala, A.; Alle, N.R.; Bellamkonda, M.K.; Dey, D.; Ganesan, S.; Biswas, M.; Sarkar, K.;
et al. Prediabetes: An overlooked risk factor for major adverse cardiac and cerebrovascular events in atrial fibrillation patients.
World J. Diabetes 2024, 15, 24–33. [CrossRef]
5. Lotfy, M.; Adeghate, J.; Kalasz, H.; Singh, J.; Adeghate, E. Chronic Complications of Diabetes Mellitus: A Mini Review. Curr.
Diabetes Rev. 2016, 13, 3–10. [CrossRef] [PubMed]
6. Kim, M.J.; Cho, Y.K.; Jung, C.H.; Lee, W.J. Association between cardiovascular disease risk and incident type 2 diabetes mellitus
in individuals with prediabetes: A retrospective cohort study. Diabetes Res. Clin. Pract. 2024, 208, 111125. [CrossRef] [PubMed]
7. Barrett-Connor, E.; Wingard, D.; Wong, N.; Goldberg, R. Heart Disease and Diabetes. In Diabetes in America, 3rd ed.; Cowie, C.,
Casagrande, S., Menke, A., Cissell, M., Eberhardt, M., Meigs, J., Gregg, E., Knowler, W., Barrett-Connor, E., Becker, D., et al., Eds.;
National Institute of Diabetes and Digestive and Kidney Diseases (US): Bethesda, MD, USA, 2018; Chapter 18.
8. Pikula, A.; Howard, B.; Seshadri, S. Stroke and Diabetes. In Diabetes in America, 3rd ed.; Cowie, C., Casagrande, S., Menke, A.,
Cissell, M., Eberhardt, M., Meigs, J., Gregg, E., Knowler, W., Barrett-Connor, E., Becker, D., et al., Eds.; National Institute of
Diabetes and Digestive and Kidney Diseases (US): Bethesda, MD, USA, 2018; Chapter 19.
9. Nanayakkara, N.; Curtis, A.J.; Heritier, S.; Gadowski, A.M.; Pavkov, M.E.; Kenealy, T.; Owens, D.R.; Thomas, R.L.; Song, S.; Wong,
J.; et al. Impact of age at type 2 diabetes mellitus diagnosis on mortality and vascular complications: Systematic review and
meta-analyses. Diabetologia 2020, 64, 275–287. [CrossRef] [PubMed]
10. Moonesinghe, R.; Yang, Q.; Zhang, Z.; Khoury, M.J. Prevalence and Cardiovascular Health Impact of Family History of Premature
Heart Disease in the United States: Analysis of the National Health and Nutrition Examination Survey, 2007–2014. J. Am. Heart
Assoc. 2019, 8, e012364. [CrossRef] [PubMed]
Appl. Sci. 2024, 14, 2132 22 of 23

11. Wong, N.D.; Zhao, Y.; Patel, R.; Patao, C.; Malik, S.; Bertoni, A.G.; Correa, A.; Folsom, A.R.; Kachroo, S.; Mukherjee, J.; et al.
Cardiovascular Risk Factor Targets and Cardiovascular Disease Event Risk in Diabetes: A Pooling Project of the Atherosclerosis
Risk in Communities Study, Multi-Ethnic Study of Atherosclerosis, and Jackson Heart Study. Diabetes Care 2016, 39, 668–676.
[CrossRef]
12. Holman, R.R.; Paul, S.K.; Bethel, M.A.; Matthews, D.R.; Neil, H.A.W. 10-Year Follow-up of Intensive Glucose Control in Type 2
Diabetes. N. Engl. J. Med. 2008, 359, 1577–1589. [CrossRef]
13. Hayward, R.A.; Reaven, P.D.; Wiitala, W.L.; Bahn, G.D.; Reda, D.J.; Ge, L.; McCarren, M.; Duckworth, W.C.; Emanuele, N.V.
Follow-up of Glycemic Control and Cardiovascular Outcomes in Type 2 Diabetes. N. Engl. J. Med. 2015, 372, 2197–2206. [CrossRef]
14. Pal, K.; Eastwood, S.V.; Michie, S.; Farmer, A.J.; Barnard, M.L.; Peacock, R.; Wood, B.; Inniss, J.D.; Murray, E. Computer-based
diabetes self-management interventions for adults with type 2 diabetes mellitus. Cochrane Database Syst. Rev. 2013, 2013, CD008776.
[CrossRef]
15. Saffari, M.; Ghanizadeh, G.; Koenig, H.G. Health education via mobile text messaging for glycemic control in adults with type 2
diabetes: A systematic review and meta-analysis. Prim. Care Diabetes 2014, 8, 275–285. [CrossRef] [PubMed]
16. Liang, X.; Wang, Q.; Yang, X.; Cao, J.; Chen, J.; Mo, X.; Huang, J.; Wang, L.; Gu, D. Effect of mobile phone intervention for
diabetes on glycaemic control: A meta-analysis: Mobile phone intervention and glycaemic control. Diabet. Med. 2011, 28, 455–463.
[CrossRef] [PubMed]
17. Li, C.; Sadraie, B.; Steckhan, N.; Kessler, C.; Stange, R.; Jeitler, M.; Michalsen, A. Effects of A One-week Fasting Therapy in Patients
with Type-2 Diabetes Mellitus and Metabolic Syndrome—A Randomized Controlled Explorative Study. Exp. Clin. Endocrinol.
Diabetes 2017, 125, 618–624. [CrossRef]
18. American Diabetes Association. Reduction in Weight and Cardiovascular Disease Risk Factors in Individuals With Type 2
Diabetes. Diabetes Care 2007, 30, 1374–1383. [CrossRef]
19. Ismail, K.; Winkley, K.; Rabe-Hesketh, S. Systematic review and meta-analysis of randomised controlled trials of psychological
interventions to improve glycaemic control in patients with type 2 diabetes. Lancet 2004, 363, 1589–1597. [CrossRef]
20. Umpierre, D. Physical Activity Advice Only or Structured Exercise Training and Association With HbA1cLevels in Type 2
Diabetes: A Systematic Review and Meta-analysis. JAMA 2011, 305, 1790. [CrossRef] [PubMed]
21. Jeon, C.Y.; Lokken, R.P.; Hu, F.B.; van Dam, R.M. Physical Activity of Moderate Intensity and Risk of Type 2 Diabetes. Diabetes
Care 2007, 30, 744–752. [CrossRef]
22. Rubin, D.J. Hospital Readmission of Patients with Diabetes. Curr. Diabetes Rep. 2015, 15, 17. [CrossRef]
23. Soh, J.G.S.; Wong, W.P.; Mukhopadhyay, A.; Quek, S.C.; Tai, B.C. Predictors of 30-day unplanned hospital readmission among
adult patients with diabetes mellitus: A systematic review with meta-analysis. BMJ Open Diabetes Res. Care 2020, 8, e001227.
[CrossRef]
24. Ogunpola, A.; Saeed, F.; Basurra, S.; Albarrak, A.M.; Qasem, S.N. Machine Learning-Based Predictive Models for Detection of
Cardiovascular Diseases. Diagnostics 2024, 14, 144. [CrossRef]
25. Joshi, R.D.; Dhakal, C.K. Predicting Type 2 Diabetes Using Logistic Regression and Machine Learning Approaches. Int. J. Environ.
Res. Public Health 2021, 18, 7346. [CrossRef] [PubMed]
26. Razavian, N.; Blecker, S.; Schmidt, A.M.; Smith-McLallen, A.; Nigam, S.; Sontag, D. Population-Level Prediction of Type 2
Diabetes From Claims Data and Analysis of Risk Factors. Big Data 2015, 3, 277–287. [CrossRef]
27. Maniruzzaman, M.; Rahman, M.J.; Ahammed, B.; Abedin, M.M. Classification and prediction of diabetes disease using machine
learning paradigm. Health Inf. Sci. Syst. 2020, 8, 1–14. [CrossRef] [PubMed]
28. Oikonomou, E.K.; Khera, R. Machine learning in precision diabetes care and cardiovascular risk prediction. Cardiovasc. Diabetol.
2023, 22, 259. [CrossRef]
29. Dagliati, A.; Marini, S.; Sacchi, L.; Cogni, G.; Teliti, M.; Tibollo, V.; De Cata, P.; Chiovato, L.; Bellazzi, R. Machine Learning
Methods to Predict Diabetes Complications. J. Diabetes Sci. Technol. 2017, 12, 295–302. [CrossRef] [PubMed]
30. Teboul, A. Diabetes Health Indicators Dataset. Kaggle, 2021. Available online: https://www.kaggle.com/datasets/julnazz/
diabetes-health-indicators-dataset (accessed on 27 January 2024).
31. Whig, P.; Gupta, K.; Jiwani, N.; Jupalle, H.; Kouser, S.; Alam, N. A novel method for diabetes classification and prediction with
PyCaret. Microsyst. Technol. 2023, 29, 1479–1487. [CrossRef]
32. Kha, Q.H.; Le, V.H.; Hung, T.N.K.; Nguyen, N.T.K.; Le, N.Q.K. Development and Validation of an Explainable Machine
Learning-Based Prediction Model for Drug–Food Interactions from Chemical Structures. Sensors 2023, 23, 3962. [CrossRef]
33. Yuan, Q.; Chen, K.; Yu, Y.; Le, N.Q.K.; Chua, M.C.H. Prediction of anticancer peptides based on an ensemble model of deep
learning and machine learning using ordinal positional encoding. Briefings Bioinform. 2023, 24, bbac630. [CrossRef]
34. Abraham, A.; Jose, R.; Ahmad, J.; Joshi, J.; Jacob, T.; Khalid, A.U.R.; Ali, H.; Patel, P.; Singh, J.; Toma, M. Comparative Analysis
of Machine Learning Models for Image Detection of Colonic Polyps vs. Resected Polyps. J. Imaging 2023, 9, 215. [CrossRef]
[PubMed]
35. Dane, S. Behavioral Risk Factor Surveillance System from Centers for Disease Control and Prevention. Kaggle, 2017. Available
online: https://www.kaggle.com/datasets/cdc/behavioral-risk-factor-surveillance-system (accessed on 27 January 2024).
36. John Clore, K.C. Diabetes 130-US Hospitals for Years 1999–2008. 2014. Available online: https://archive.ics.uci.edu/dataset/29
6/diabetes+130-us+hospitals+for+years+1999-2008 (accessed on 4 February 2024).
Appl. Sci. 2024, 14, 2132 23 of 23

37. Strack, B.; DeShazo, J.P.; Gennings, C.; Olmo, J.L.; Ventura, S.; Cios, K.J.; Clore, J.N. Impact of HbA1c Measurement on Hospital
Readmission Rates: Analysis of 70,000 Clinical Database Patient Records. BioMed Res. Int. 2014, 2014, 781670. [CrossRef]
[PubMed]
38. Shariq, O.A.; McKenzie, T.J. Obesity-related hypertension: A review of pathophysiology, management, and the role of metabolic
surgery. Gland Surg. 2020, 9, 80–93. [CrossRef] [PubMed]
39. Pakhare, M.; Anjankar, A. Critical Correlation Between Obesity and Cardiovascular Diseases and Recent Advancements in
Obesity. Cureus 2024, 16, e51681. [CrossRef] [PubMed]
40. Yan, Z.; Cai, M.; Han, X.; Chen, Q.; Lu, H. The Interaction Between Age and Risk Factors for Diabetes and Prediabetes: A
Community-Based Cross-Sectional Study. Diabetes Metab. Syndr. Obes. 2023, 16, 85–93. [CrossRef]
41. Besseling, J.; Kastelein, J.J.P.; Defesche, J.C.; Hutten, B.A.; Hovingh, G.K. Association Between Familial Hypercholesterolemia and
Prevalence of Type 2 Diabetes Mellitus. JAMA 2015, 313, 1029. [CrossRef]
42. Toma, M.; Wei, O.C. Predictive Modeling in Medicine. Encyclopedia 2023, 3, 590–601. [CrossRef]
43. Bekbolatova, M.; Mayer, J.; Ong, C.W.; Toma, M. Transformative Potential of AI in Healthcare: Definitions, Applications, and
Navigating the Ethical Landscape and Public Perspectives. Healthcare 2024, 12, 125. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.