Deep Learning Models for Type 2 Diabetes Detection in Saudi Arabia

Insulin resistance and deficiency are the main causes of type 2 diabetes (T2D), which raises blood sugar levels [1]. The illness has spread around the world in recent decades and impacted a large number of people [2]. According to estimates from the World Health Organisation (WHO), 3 million Saudi Arabians have prediabetes, or the risk of developing diabetes. According to WHO data, 7 million Saudi Arabians suffer from diabetes, which has a serious impact on their lives [3]. Furthermore, with an estimated 24% of the adult population suffering from T2D, Saudi Arabia is rated seventh among nations with a high prevalence of T2D [4], [5]. Factors including urbanisation, obesity rates rising quickly in the area, and sedentary lifestyles are to blame for the large number of patients [5]. Early detection and the right kind of intervention can help manage T2D. This will postpone or avoid the disease's consequences, like retinopathy, heart problems, and renal failure. Early identification hence improves treatment outcomes and lessens the load on healthcare systems [6]. Healthcare problems including diagnosis and disease identification have improved thanks to the inclusiveness of deep learning and machine learning approaches [7]. Furthermore, the methods' computation has affected the early detection of T2D; for this reason, analyzing the medical records is necessary. However, a number of studies and papers have been published on the superiority of deep learning and machine learning models over antiquated statistical techniques in indicating the risk of diabetes [8].

Motivation. Because it aids in the management and prevention of issues like nephropathy, retinopathy, and cardiovascular disease, it is crucial. Certain conventional screening techniques, such blood glucose testing and oral glucose tolerance tests, take a lot of time and don't always yield reliable results [9]. Thus, there has been interest in creating a disease detection technique that is accurate, dependable, and economical. Due to their ability to analyse massive datasets, machine learning and deep learning techniques have been shown to be effective [7]. Because the approaches analyse complex data automatically rather than by hand, they are more efficient than standard statistical methods and have been applied in a variety of healthcare applications [10]. Numerous studies have looked into the use of deep learning and machine learning methods for T2D prediction in recent years. Deep learning algorithms are being used to diagnose type 2 diabetes in Saudi population, although there is a dearth of research on this topic. There aren't many datasets available in related literature for Saudi population T2D detection. One of these datasets was gathered [3], and the dataset sub-section of this work will detail it. Despite this, the authors in [3] obtained the greatest accuracy of 78.9% using a decision forest classifier for unbalanced data and 82.1% with balanced data using the Synthetic Minority Over-sampling Technique (SMOTE). They did, however, only employ a few machine learning algorithms. Furthermore, comparable research regarding the application of deep learning models to T2D datasets is currently lacking.

Deep Learning Applications

1. Image Recognition: Deep learning models, particularly Convolutional Neural Networks (CNNs), have excelled in image classification and object detection tasks [11].
2. Natural Language Processing (NLP): Models like BERT have set new benchmarks in a variety of NLP tasks, including question answering and sentiment analysis [12].
3. Autonomous Vehicles: Deep learning algorithms are vital for interpreting sensory input and making driving decisions in self-driving cars [13].
4. Game Playing: Deep reinforcement learning models, such as AlphaGo, have defeated world champions in complex games like Go [14].
5. Healthcare and Medicine: Deep learning is utilized in predictive analytics, disease identification, and drug discovery [15].
6. Speech Recognition: Deep learning models are extensively used in developing accurate speech recognition systems [16].
7. Fraud Detection: Deep learning assists in identifying fraudulent activities through pattern recognition in transaction data [17].
8. Customer Service: Chatbots and virtual assistants employing deep learning offer enhanced customer service by understanding and responding to user queries effectively [18].
9. Facial Recognition: Deep learning is used in facial recognition systems for security and authentication purposes [19].

Contributions. There are two gaps related to Type 2 Diabetes (T2D) detection in Saudi Arabia that this research aims to fill to contribute to the field. Some of the ways that this study will contribute to the field include:

1. The development of deep learning models: In this study, various deep learning models are tailored for T2D detection in the Saudi population, which are developed and implemented. These deep learning models are Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Autoencoder (AE). All these three models are advanced computational techniques leveraged in the study to determine how effective they are in identifying complex medical patterns among T2D patients. It is important to note that previous Saudi Arabian studies in literature have failed to explore this area thoroughly.
2. Comparative performance analysis: This paper further contributes to the existing literature by comparing the deep learning models that we developed with Decision Forest, a traditional machine learning classifier. The DNN, AE, and CNN comparisons with Decision Forest were performed using a Saudi-specific dataset, which has been referenced [3]. These comparisons were done with and without the Synthetic Minority Oversampling Technique (SMOTE) application, which is essential for addressing class imbalance. The comparisons of this study were made up of two scenarios, which used available features and reduced six out of ten features, which is an approach similar to the one taken in the Decision Forest [3].

The datasets used in this work were both balanced and imbalanced data. These datasets are essential because they provide important information regarding data preprocessing techniques such as SMOTE, which influence the performance of deep learning and traditional models. This research examined the performances of both a complete feature set and a reduced one, which enabled we to evaluate their capabilities and techniques that can be adopted to optimize them for detecting T2D in Saudi Arabia. As a result, this study contributed innovatively to how the current gaps in literature can be bridged by introducing deep learning methods to the Saudi T2D detection landscape and proposing a rigorous comparative framework, which can be used by future studies and in clinical decision-making processes. The results of this study can be used to enhance both T2D screen and diagnosis effectiveness and precision in Saudi Arabia.

This paper is structured into distinct sections: Section 2 delineates the models employed for diabetes detection. Section 3 presents materials and methods used. Section 4 exhibits the experimental results, and Section 5 provides a conclusive summary.

The methods used to detect diabetes, particularly type 2 diabetes, are displayed in this section. With the aid of machine learning techniques, Farooq Ahmad et al. [20] elaborates on the factors that predict T2D. Three thousand patients' worth of records from several Saudi hospitals were used by the researchers. Afterwards, they used preprocessing methods and discussed their importance, resulting in a 162 case decrease. Modelling using the techniques of ensemble majority voting (EMV), random forest (RF), logistic regression (LR), decision tree (DT), and support vector machine (SVM). SVM obtained 82.10% in the first dataset, RF with nine features achieved 88.27% accuracy, and RF with eight features obtained 87.65% accuracy in the second dataset. Additionally, Syed and Khan [3] developed an application that might be utilised to identify T2D in Saudi Arabia. Data for their study were obtained from King Abdulaziz University and included 3906 non-diabetic subjects and 990 diabetic cases. To identify important features, binary LR and the Pearson chi-squared test were employed. The dataset underwent preprocessing using an 80:20 ratio, dividing it into training and testing sets. The Synthetic Minority Over-sampling Technique (SMOTE) was used to attain equilibrium. After utilising nine different binary classification methods, it was discovered that the Decision Forest (DF) approach outperformed other models.

Gollapalli et al. [4] created an efficient model for diabetes that was able to predict and detect three types of diabetes: Type 1 Diabetes, T2D, and prediabetic. They did this by applying a variety of machine learning classifiers on a dataset that was obtained from King Fahad University Hospital (KFUH), a hospital in Saudi Arabia. There were 897 instances and 10 different attributes in the sample. The writers or researchers used stacking techniques, SVM, Bagging, DT, K-nearest neighbour (KNN), and RF as their main classifiers. To maximise the outcomes, four trials were carried out, with experiments 2, 3, and 4 employing SMOTE to balance the dataset. Their novel stacking model combined KNN with a KNN meta-classifier, Bagging DT, and Bagging KNN to obtain an accuracy of 94.48%. Sex, education, antiDiab, and nutrition were the five factors that were found to have a significant impact on the accuracy of the model through a critical study of feature importance. Alassaf et al. [21] presented a method intended to proactively diagnose diabetes in an uncharted area. They received data from Khobar's KFUH, which was the first time the information was used to support a diagnosis. The authors performed pre-processing and identified key traits before classifying the data. In addition, they employed recursive feature elimination and the correlation coefficient for feature selection. Following that, four categories of algorithms—Naïve Bayes (NB), Artificial Neural Networks (ANN), Support Vector Machines (SVM), and KNN—were evaluated with an emphasis on classification accuracy, F-measure, precision, and recall. ANN fared better than the other models, with 77.5% accuracy.

Alanazi and Mezher proposed a model combining RF and SVM classifiers to predict diabetes. They obtained a real dataset from the primary health care unit of the security forces in Tabuk, Saudi Arabia. Their model employed RF showed a remarkable performance with an accuracy of 98% and 99% receiver operating characteristic curve (ROC) boast. It signifies that the RF method was much better than SVM in accuracy [22]. In [23], the research project leveraged real healthcare datasets comprising 18 attributes, sourced from the Ministry of National Guard Health Affairs (MNGHA) database. The primary objective was to construct a predictive model for identifying diabetic patients within the adult population of Saudi Arabia. Three distinct algorithms, namely the Self-Organizing Map (SOM), C4.5, and RF, were applied for this purpose. Comparative analysis against various classifiers revealed that RF consistently delivered superior performance.

In the investigation of T2D, Jaber and James in [24] employed diverse classifiers, including the NB Algorithm, the LR Algorithm, and the RF Algorithm. The Pima Indian Diabetes Dataset (PIDD) served as the foundational dataset. The findings underscored the supremacy of the RF Algorithm when compared to alternative approaches. Various machine learning algorithms including linear discriminate (LD), linear SVM, quadratic SVM, cubic SVM, Gaussian SVM, fine KNN, weighted KNN and neural pattern recognition (NPR) were harnessed to construct classification models for diabetes detection in [25]. The result of a rigorous performance analysis showed that the weighted KNN showed commendable predictive accuracy in estimating the prevalence of diabetes in both male and female datasets, with an impressive average accuracy of 94.5% and shorter training time compared to other classification methods.

The identified areas for improvement identified in the related work with this study that could lead to improvements included:

• Limited Deep Learning Exploration: Most of the studies that have been conducted in the literature are those that provide insights regarding machine learning and T2D. The literature analysis indicates a massive gap regarding the impact of deep learning methods in Saudi Arabia. Common deep learning models overlooked in studies conducted in Saudi Arabia include DNN, AE, and CNN, which tend to have significant success in domains such as image recognition, natural language processing, and bioinformatics. DNN, AE, and CNN are deep learning methods that can assist in exposing intricate patterns in complex T2D detection medical data. However, the sad reality is that these deep learning methods in Saudi Arabia are untapped areas.
• Absence of Comparative Research: Currently, no comparative studies have been conducted in Saudi Arabia using the same dataset used to compare deep learning methods and the traditional machine learning models for T2D detection. This is an issue since comparative studies are vital for a better understanding of the relative advantages and limitations of various analytical approaches. The lack of such studies makes it extremely difficult to conclude whether deep learning methods can help address machine learning models' accuracy, speed, and diagnosis weaknesses in detecting T2D.
• Unavailability of Public and Free Datasets for T2D Detection in Saudi Arabia: The lack of available and free dataset within the Saudi population, that is accessible to researchers, cause a substantial hiatus in the present research landscape for T2D. The ability of the researchers to develop models and carry out thorough studies that can be generalized for the entire population is limited, because the datasets available are usually kept within the hospitals and not generally released to the public. This in turn obstructs the progress by restricting the extent of possible research and impedes the development of strong diagnostic tools that can benefit the larger community. This problem in Saudi Arabia could be bridged by establishing an open-access, anonymized dataset, potentially, catalyzing advancements in detecting and managing T2D.

This study made use of deep learning approaches to compare T2D classifications. The two cases that this study considered are those that use all features and those that used 6 out of 10 features as used with the Decision Forest in [3]. Both cases use balanced data (with SMOTE) and imbalanced data (without SMOTE). The whole system for detecting the T2D disease is depicted in Figure 1. The dataset is first preprocessed. The preprocessed dataset was then separated into two pieces: (1) training set; (2) testing set. The system uses different deep learning models such as DNN, AE, and CNN. Finally, a set of performance measurement metrics are used to assess each model’s effectiveness such as accuracy, precision, recall, F1-score, and area under the ROC curve (AUC), in which ROC acts as the receiver operating characteristic curve.

A) Dataset

The dataset used was collected by Syed and Khan in [3]. This dataset is named Saudi Dataset (SD) to be easier to follow. SD is collected by filling out a cross-sectional survey. Participants from the King Abdulaziz University (KAU) were issued with a fill-in form to complete. The survey utilized close-ended questions to gather data regarding diabetes risk factors for the participants from KAU to predict the occurrence of diabetes in Saudi Arabia’s western part. The researchers extracted the most common attributes of diabetes prediction from models of diabetes papers that had recently been published when preparing the survey. The researchers used non-invasive tests and direct observation techniques to address attributes in the survey. However, the researchers first obtained the necessary permissions from the KAU Deanship of Graduate Studies before conducting the study at the university. The study participants from KAU were students, staff, and faculty members. The survey contains eleven questions as follows:

1. Region.
2. Age group.
3. Sex.
4. Body Mass Index (BMI).
5. Waist size (navel level measurement).
6. Do you do a minimum of 30 minutes of physical activity every day?
7. Fruits and vegetables daily intake.
8. Are you currently undergoing hypertension treatment?
9. Does your family tree have a history of diabetes?
10. Do you smoke?
11. Has your blood glucose been high for any reasons (such as pregnancy, illness, etc.)?

The total number of subjects in this study was 4896, of which 990 were those at high risk, and the remaining 3906 were those at low risk of having diabetic complications. The diabetic data attributes of the participants are shown in Table 1, in which the researchers used the Label Encoding method as a preprocessing step to code the attributes. Q1-Q10 represent either the explanatory variables or predictors of this study. We used the survey's last question as a categorical response variable retrieving data on the level of High Fasting Blood Glucose. It was agreed that those who answered “Yes” to the categorical response variable question would be considered that they were at a high risk of developing diabetes. For those who answered "No" to the categorical response variable question would be considered that they were at a low risk of developing diabetes [3].

B) Deep Learning Models

This part defines the used deep learning models including DNN, AE, and CNN.

• Deep Neural Networks (DNN): Deep Neural Networks are neural architectures where neurons are structured in a connected series across multiple layers. Neurons in each layer receive activations from the previous layer, creating a continuous series of interconnected neurons. These neurons collectively perform computations on input data, which include weighted summation followed by nonlinear activation functions. Consequently, DNN exhibit complex and nonlinear mappings from input to output. Through the backpropagation technique, DNN have the capacity to learn these intricate mappings from data by adjusting the weights of each neuron [28].
• Autoencoder (AE): The autoencoder (AE), also referred to as an autoassociator, belongs to the family of Artificial Neural Networks (ANN) and operates as an unsupervised learning algorithm. Its primary objective is to encode datasets, effectively reducing dimensions. Historically, autoassociators have been the focus of extensive research within the field of ANN. Recently, autoencoders have found significant application in the realm of learning generative data models. The typical workflow begins with inputting data, which is subsequently transformed into an abstract representation. Following this, the encoder mechanism comes into play, converting the abstract representation back to its original format. The encoder possesses the capacity to encode the input into a distinct representation, with the ultimate goal of ensuring seamless reconstruction of the input data from that representation. Throughout this process, the autoencoder diligently endeavors to establish an identity function. A noteworthy attribute of autoencoders lies in their ability, during the propagation phase, to systematically eliminate extraneous data features while retaining valuable information. This coding process frequently involves transforming the input vector into a lower-dimensional representation, thereby enhancing the overall efficiency of the learning process [29], [30].
• Convolutional Neural Networks (CNN): Commonly known as CNN, Convolutional Neural Networks are neural architectures organized hierarchically, featuring layers that alternate with subsampling layers. These neural networks draw inspiration from the simple and complex cells found in the human visual cortex. Collectively, these hierarchical neural networks comprise fully connected layers analogous to multilayer perceptrons (MLP). CNN operate in a manner akin to the human visual system, adept at identifying patterns and structures in visual data. Since their inception, CNN have gained substantial popularity and have become one of the prime choices for a wide range of deep learning tasks, including object recognition in large images [31]. In this paper, CNN with one dimension is used.

The algorithms were executed using Python programming language. They can be run on Google Colab, providing it with access to GPUs, which help to speed up the training process significantly. When the algorithm is run on Google Colab, it freely accesses the GPUs, developing and running the Python code directly without requesting additional configurations [35]. At the backend, TensorFlow is used to build and train the programs' networks written in Python 3.<\p>

A) Performance Measures

Evaluation of the performance of the proposed models was done using various measurements such as Accuracy (Acc), precision, recall, F1-measure, and area under the ROC curve (AUC), in which ROC acts as the receiver operating characteristic curve. The model's accuracy is defined as the patient proportion that the models diagnosed properly. The accuracy of the model’s formula is shown in below [36]: \begin{equation}\label{e1} Accuracy (Acc)=\frac{TN+TP}{TN+TP+FN+FP} \end{equation}

True positive (TP) is a term used to refer to the population of patients classified as positive and are truly positive. True negative (TN) represents the predicted patients that are negative and actually negative. False positive (FP) in the equation represents the patients classified as positive, while in the real sense are negative. False negative (FN) refers to the number of patients that are categorized as negative but are actually positive. To determine the proposed model’s classification quality, the parameters of the access are regularly estimated [36]. Precision is the second performance evaluation metric that refers to the sum of true positive and true negative. The formula for calculating precision is presented in below [37]: \begin{equation}\label{e2} Precision=\frac{TP}{TP+FP} \end{equation}

Recall refers to the attributes that are classified correctly and is also known as sensitivity or the True Positive Rate. Recall is expressed as the total number of positive predictions divided by the total number of class values. The formula for calculating recall is presented in below [37]: \begin{equation}\label{e3} Recall=\frac{TP}{TP+FN} \end{equation}

The F1-measure is also known as the F1-score, and it is a performance metric since it can provide data between the recall and precision. The formula used to calculate F1-measure is presented in below [37]:

\begin{equation}\label{e4} F1-Score=2 \times \frac{Precision*Recall}{Precision+Recall} \end{equation}

The Area under the Curve (AUC) value can be used to measure the discriminative power of classification algorithms. AUC is used to assess models' performance with values ranging from 0 to 1. Values of 1 or near 1 mean the model is excellent at finding the balance between recall and precision. Such values indicate models capable of producing superior classification performance [3].

B) Result

(i) Case 1: Using the entire dataset’s features

In the first case, we used the dataset's entire features as outlined in Table 1 above. This study went a step ahead to explore how three deep learning models compare with one another in detecting T2D in data that was imbalanced (does not use SMOTE) and balanced (makes use of SMOTE). The SMOTE is currently one of the methods most used to address imbalanced classes. Thus, this study made use of the SMOTE to obtain balanced data. As shown in Figure 2, the metrics that we used to evaluate the developed deep learning methods are accuracy, F1 score, precision, AUC, and recall. Figure 2 (a) below shows the imbalanced data for this study:

• DNN: This is one of the models that demonstrated good accuracy and AUC of 80.61% and 83.02%, respectively. However, it was evident that DNN had a moderate precision of 54.44% and low recall and F-1 score of 32.25% and 39.66%, respectively. These results indicate that DNN is unreliable for identifying positive cases in imbalanced data, as some challenges exist.
• AE: This model demonstrated that it could deliver superior performances in recall and F-1 scores of 46.88% and 50.05%, respectively. AE also recorded excellent performance in accuracy and AUC of 81.12% and 84.3%, respectively. The scores of this model indicate that it is reliable for identifying diabetic cases, especially in imbalanced datasets.
• CNN: The results of this model were slightly similar to DNN in specific parameters and significantly differed in others. For example, the accuracy score of this model and AUC were 80.41% and 83.02%, respectively. However, CNN had lower scores than DNN in precision, recall, and F-1, with 54.14%, 23.06%, and 32.04%, respectively. As a result, it can be concluded that the CNN is a model that is not entirely reliable in detecting diabetes cases in patients.

Based on the scores of the three models developed in this study, it can be argued that AE is the best model for detecting cases of diabetics. This is because AE has the highest score among the three models in recall, F1 score, AUC, and accuracy. Although DNN outperformed the other models, especially in precision, AE had the highest scores in all the other metrics. Figure 2 (b) shows the results of the three models for balanced data. The summary of the scores is as follows:

• DNN: This model recorded strong performances in recall of 85.03% but moderate accuracy, precision, and F-1 scores of 79.1%, 49.12%, and 62.27%. Nevertheless, the excellent recall score reveals that DNN is an effective model for identifying positive cases in balanced data.
• AE: Again, AE had some of the best performances when compared to the other two models for detecting diabetic cases in balanced data. For example, the scores of this model were: accuracy (79.16%), precision (49.22%), F1 score (62.28%), recall (84.8%), and AUC (83.7%). These scores indicate that AE is quite reliable in distinguishing diabetic classes.
• CNN: When compared generally to metrics, it shows lower scores, with accuracy at 78.94%, precision at 48.85%, recall at 80.87%, and F1-score at 60.9%. Its AUC reflects DNN (83.02%).

Again, AE is the most balanced model compared to the other models, as evidenced by its metrics scores. AE is the model that leads in most metrics, such as accuracy, precision, F1-score, and AUC. Although DNN had an excellent recall score, AE performed better in other metrics, revealing that it is more accurate and reliable for detecting people with diabetes in imbalanced datasets. Based on the results of the imbalanced and balanced datasets, it is evident that AE is the best model for identifying cases of diabetes in both of these datasets. This is because out of the three models, AE had superior performances in accuracy, F1-score, and AUC. In contrast, DNN and CNN seemed to lag, recording some of the poorest performances in recall and F1-score metrics, especially on imbalanced data.

(ii) Case 2: Using six features out of ten features

Another case that this study used to compare the three deep learning models that were developed and the Decision Forest (DF), popularly known for outperforming most machine learning algorithms in \cite{3}. All these four models were used to detect T2D for imbalanced and balanced data. The imbalanced data was the data that did not use SMOTE, while the balanced data was the data that was balanced using SMOTE. These four models for detecting diabetic cases were evaluated using five metrics, which are accuracy, AUC, precision, F1-score, and recall, as shown in Figure 3. Figure 3 (a) shows the DNN, AE, CNN, and DF results when using imbalanced dataset. The summary of the results of these models are as follows:

• DNN: This deep learning model had average scores in accuracy and AUC of 80.22% and 83.63%, respectively. However, DNN had poor scores in precision and recall, at 53.54% and 21.71%, respectively. These scores indicate that DNN has limited effectiveness in accurately identifying positive diabetes cases.
• AE: This deep learning model was the most promising as it recorded some of the highest scores on multiple metrics such as accuracy (81.01%), recall (40.71%), F1- score (46.02%), and AUC (84.97%). The scores of AE reveal that it has an excellent balanced performance, and it effectively identifies true positives in imbalanced datasets.
• CNN: This model had one of the best precision scores among the three models at 55.8%. However, CNN recorded the worst scores in recall and F1-score of 9.73% and 16.52%, respectively. The results of this model demonstrate challenges with using CNN to classify positive cases of diabetes correctly.
• DF: Amongst the models, DF displayed the least performance in AUC (82.2%), accuracy (78.9%), precision (46.4%), F1-Score (43), and recall (40%). Although there was some balance between precision and recall it mostly lagged behind the deep learning models.

Generally, the results of the four models reveal that AE is the best for detecting diabetic cases in imbalanced datasets. This is because AE had the highest scores in all the performance metrics used in this study. As a result, it can be concluded that AE is the most reliable model for identifying true positive cases of people with diabetes and balancing class differentiation. On the contrary, DNN scores indicate that although it has an overall moderate performance, it struggles a lot with recall, making it slightly unreliable in detecting positive cases. CNN is another model with varying scores across the metrics used in this study. While this model has excellent precision performance, it had some of the worst recall and F1 scores. The results of the recall and F1-score reveal that CNN is not a reliable classification model. DF was a model with balanced performances in precision and recall but failed to record promising scores in the other metrics. The performance of DF across all the metrics suggests that deep learning methods are more reliable, especially AE.

Figure 3 (b) shows the DNN, AE, CNN, and DF results when using balanced datasets. The summary of the scores of these four models, when balanced data is used, are as follows:

• DNN: This model had some of the best scores, evident in recall (86.61%) and accuracy (78.88%). However, DNN had moderate scores in precision and F1-score of 48.84% and 62.46%, respectively. These results suggest that DNN is best suited for identifying positive cases.
• AE: This model performed better than DNN in most metrics. In fact, the only metric that DNN had a better score than AE is recall, in which its performance score was 86.45%. AE's accuracy score was 78.98%, precision score was 48.79%, F1-score was 62.52%, and AUC was 84.47%.
• CNN: This model had the second-highest accuracy score of 79.01%. It also had good scores in other metrics, scoring 49.02% in precision, 86.38% in recall, and 62.54% in F1-score. The results of this model reveal that it is accurate in identifying T2D and classification.
• DF: This model outperformed all three deep learning models, recording some of the highest scores across all metrics. DF had the highest accuracy score of 82.1%, precision score of 77.6%, recall of 89%, F1-score of 82.9%, and AUC of 86.7%.

The results of these four models reveal that DF is the superior model, recording the highest accuracy, precision, recall, F1-score, and AUC scores. This shows that DF is the best model for detecting cases of diabetes in the balanced data context. Conversely, although DNN has an excellent recall score, it has moderate scores in accuracy and precision. However, while DF outperforms AE and CNN, the two deep-learning models have better scores than DNN. Compared to DNN, AE and CNN have more balanced performance profiles, making them more ideal. Overall, it can be concluded that DF is a better model for correctly identifying positive cases of diabetes and ensuring high accuracy in predictions when using balanced datasets. In analyzing the three deep learning models, AE is a better model as it had better accuracy, recall, F1-score, and AUC scores for imbalanced data. On the other hand, DF is the better model for balanced data as it performs better than the three deep learning models.

Early detection of Type 2 Diabetes (T2D) is critical for implementing appropriate treatment strategies and lifestyle adjustments. These will help decelerate and prevent the advancement of the condition. Studies in machine learning and medical datasets have overlooked the Saudi population on T2D detection. This study addresses this research gap by developing and comparing three deep learning methods and the traditional classifier. The deep learning methods developed are Deep Neural Network (DNN), Autoencoder (AE), and Convolutional Neural Network (CNN), while the traditional classifiers used are the Decision Forest (DF). As a result, this study contributes to the literature by providing a comparative analysis of different models, an aspect that was missing. The performance metrics used to evaluate these four models are accuracy, precision, recall, F1-score, and AUC. The findings of this study revealed that AE outperformed other models in the case of using all features of the dataset for imbalanced and balanced data with the highest accuracy of 81.12% and 79.16%, respectively. For the case of using 6 features, the results showed that AE achieved the highest accuracy of 81.01% with imbalanced data compared to DF and DF achieved the highest accuracy of 82.1% with balanced data. Findings suggest that AE is superior when imbalanced and balanced datasets with all features are used. However, DF is a more reliable model when using balanced data with fewer features. Overall, AE is more useful in detecting T2D cases that are yet to be detected. These results signify that the model can be a valuable tool for healthcare practitioners in Saudi Arabia. The models developed in this research may be employed as a screening instrument in public health initiatives. It will also support awareness goals about T2D. This in turn will promote healthier lifestyles among different populations. Future investigations should be expanded by incorporating a larger and more diverse dataset. This should encompass additional variables like genetic factors and dietary patterns. Alternative deep learning architectures and integrating advanced techniques such as data augmentation and transfer learning will lead to improved results. Future research studies will help authenticate the models using more extensive datasets and assess their practical viability within clinical contexts.

The authors declare no conflict of interests. All authors read and approved final version of the paper.

All authors contributed equally in this paper.

1. American Diabetes Association. (2019). Standards of Medical Care in Diabetes-2019 Abridged for Primary Care Providers. Clinical Diabetes, 37(1), 11-34.
2. Saeedi, P., Petersohn, I., Salpea, P., Malanda, B., Karuranga, S., Unwin, N., ... & IDF Diabetes Atlas Committee. (2019). Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: Results from the International Diabetes Federation Diabetes Atlas. Diabetes Research and Clinical Practice, 157, 107843.
3. Syed, A. H., & Khan, T. (2020). Machine learning-based application for predicting risk of type 2 diabetes mellitus (t2dm) in saudi arabia: A retrospective cross-sectional study. IEEE Access, 8, 199539-199561.
4. Gollapalli, M., Alansari, A., Alkhorasani, H., Alsubaii, M., Sakloua, R., Alzahrani, R., ... & Albaker, W. (2022). A novel stacking ensemble for detecting three types of diabetes mellitus using a Saudi Arabian dataset: Pre-diabetes, T1DM, and T2DM. Computers in Biology and Medicine, 147, 105757.
5. Alwin Robert, A., Abdulaziz Al Dawish, M., Braham, R., Ali Musallam, M., Abdullah Al Hayek, A., & Hazza Al Kahtany, N. (2017). Type 2 diabetes mellitus in Saudi Arabia: major challenges and possible solutions. Current Diabetes Reviews, 13(1), 59-64.
6. Tabák, A. G., Herder, C., Rathmann, W., Brunner, E. J., & Kivimäki, M. (2012). Prediabetes: a high-risk state for developing diabetes. Lancet, 379(9833), 2279-2290.
7. Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347-1358.
8. Krittanawong, C., Zhang, H., Wang, Z., Aydar, M., & Kitai, T. (2017). Artificial intelligence in precision cardiovascular medicine. Journal of the American College of Cardiology, 69(21), 2657-2664.
9. American Diabetes Association. (2020). 2. Classification and Diagnosis of Diabetes: Standards of Medical Care in Diabetes—2020. Diabetes Care, 43(Supplement 1), S14-S31.
10. Kavakiotis, I., Tsave, O., Salifoglou, A., Maglaveras, N., Vlahavas, I., & Chouvarda, I. (2017). Machine learning and data mining methods in diabetes research. Computational and Structural Biotechnology Journal, 15, 104-116.
11. Li, Y. (2022, January). Research and application of deep learning in image recognition. In 2022 IEEE 2nd international conference on Power, Electronics and Computer Applications (ICPECA) (pp. 994-999). IEEE.
12. Otter, D. W., Medina, J. R., & Kalita, J. K. (2020). A survey of the usages of deep learning for natural language processing. Ieee Transactions on Neural Networks and Learning Systems, 32(2), 604-624.
13. Miglani, A., & Kumar, N. (2019). Deep learning models for traffic flow prediction in autonomous vehicles: A review, solutions, and challenges. Vehicular Communications, 20, 100184.
14. Justesen, N., Bontrager, P., Togelius, J., & Risi, S. (2019). Deep learning for video game playing. IEEE Transactions on Games, 12(1), 1-20.
15. Mittal, S., & Hasija, Y. (2020). Applications of deep learning in healthcare and biomedicine. Deep Learning Techniques for Biomedical and Health Informatics, 57-77.
16. Lee, W., Seong, J. J., Ozlu, B., Shim, B. S., Marakhimov, A., & Lee, S. (2021). Biosignal sensors and deep learning-based speech recognition: A review. Sensors, 21(4), 1399.
17. Roy, A., Sun, J., Mahoney, R., Alonzi, L., Adams, S., & Beling, P. (2018, April). Deep learning detecting fraud in credit card transactions. In 2018 Systems and Information Engineering Design Symposium (Sieds) (pp. 129-134). IEEE.
18. Nuruzzaman, M., & Hussain, O. K. (2018, October). A survey on chatbot implementation in the customer service industry through deep neural networks. In 2018 IEEE 15th International Conference on e-Business Engineering (ICEBE) (pp. 54-61). IEEE.
19. Ríos-Sánchez, B., Costa-da-Silva, D., Martín-Yuste, N., & Sánchez-Ávila, C. (2019). Deep learning for facial recognition on single sample per person scenarios with varied capturing conditions. Applied Sciences, 9(24), 5474.
20. Ahmad, H. F., Mukhtar, H., Alaqail, H., Seliaman, M., & Alhumam, A. (2021). Investigating health-related features and their impact on the prediction of diabetes using machine learning. Applied Sciences, 11(3), 1173.
21. Alassaf, R. A., Alsulaim, K. A., Alroomi, N. Y., Alsharif, N. S., Aljubeir, M. F., Olatunji, S. O., ... & Alturayeif, N. S. (2018, April). Preemptive diagnosis of diabetes mellitus using machine learning. In 2018 21st Saudi Computer Society National Computer Conference (NCC) (pp. 1-5). IEEE.
22. Alanazi, A. S., & Mezher, M. A. (2020, September). Using machine learning algorithms for the prediction of diabetes mellitus. In 2020 International Conference on Computing and Information Technology (ICCIT-1441) (pp. 1-3). IEEE.
23. Daghistani, T., & Alshammari, R. (2016). Diagnosis of diabetes by applying data mining classification techniques. International Journal of Advanced Computer Science and Applications, 7(7), 329-332.
24. Jaber, F. A., & James, J. W. (2023). Early prediction of diabetic using data mining. SN Computer Science, 4(2), 169.
25. Almutairi, E. S., & Abbod, M. F. (2023). Machine learning methods for Diabetes prevalence classification in Saudi Arabia. Modelling, 4(1), 37-55.
26. Alam, T. M., Iqbal, M. A., Ali, Y., Wahab, A., Ijaz, S., Baig, T. I., ... & Abbas, Z. (2019). A model for the early prediction of diabetes. Informatics in Medicine Unlocked, 16, 100204.
27. Sneha, N., & Gangil, T. (2019). Analysis of diabetes mellitus for early prediction using optimal features selection. Journal of Big data, 6(1), 1-19.
28. Montavon, G., Samek, W., & Müller, K. R. (2018). Methods for interpreting and understanding deep neural networks. Digital Signal Processing, 73, 1-15.
29. Pathirage, C. S. N., Li, J., Li, L., Hao, H., Liu, W., & Ni, P. (2018). Structural damage identification based on autoencoder neural networks and deep learning. Engineering Structures, 172, 13-28.
30. Liu, W., Wang, Z., Liu, X., Zeng, N., Liu, Y., & Alsaadi, F. E. (2017). A survey of deep neural network architectures and their applications. Neurocomputing, 234, 11-26.
31. Kiranyaz, S., Ince, T., & Gabbouj, M. (2015). Real-time patient-specific ECG classification by 1-D convolutional neural networks. IEEE Transactions on Biomedical Engineering, 63(3), 664-675.
32. Muni Kumar, N., & Manjula, R. (2019). Design of multi-layer perceptron for the diagnosis of diabetes mellitus using keras in deep learning. In Smart Intelligent Computing and Applications: Proceedings of the Second International Conference on SCI 2018, Volume 1 (pp. 703-711). Springer Singapore.
33. Zhou, H., Myrzashova, R., & Zheng, R. (2020). Diabetes prediction model based on an enhanced deep neural network. EURASIP Journal on Wireless Communications and Networking, 2020, 1-13.
34. Tabares-Soto, R., Orozco-Arias, S., Romero-Cano, V., Bucheli, V. S., Rodríguez-Sotelo, J. L., & Jiménez-Varón, C. F. (2020). A comparative study of machine learning and deep learning algorithms to classify cancer types based on microarray gene expression data. PeerJ Computer Science, 6, e270.
35. Pandiya, M., Dassani, S., & Mangalraj, P. (2020). Analysis of deep learning architectures for object detection-a critical review. 2020 IEEE-HYDCON, 1-6.
36. Rahman, M., Islam, D., Mukti, R. J., & Saha, I. (2020). A deep learning approach based on convolutional LSTM for detecting diabetes. Computational Biology and Chemistry, 88, 107329.
37. Naz, H., & Ahuja, S. (2020). Deep learning approach for diabetes prediction using PIMA Indian dataset. Journal of Diabetes & Metabolic Disorders, 19, 391-403.