Diabetes mellitus (DM) is a chronic metabolic disease which affects the human body by impairing its ability to convert blood sugar to energy (1). People diagnosed with diabetes are unable to regulate their blood sugar level in the body, which resulted in high levels of blood pressure and sugar (2). If diabetes is not detected, diagnosed, and treated in the early phases, it causes life-threatening diseases like kidney failure, diabetic retinopathy, and various cardiovascular diseases (3). Despite the advances in the medical field in recent times, diabetes remains challenging in many societies (4). Therefore, it is crucial to design intelligent systems which provide support to medical personnel in the diagnosis of diabetes and decision-making (5). The classical lab tests that depend on diabetic diagnosis approaches are expensive and time-consuming (6). Generally, clinicians consider an approximate diagnosis and prediction of diabetes mellitus by considering fasting blood sugar or random blood sugar tests (7). The glycated hemoglobin (HbA1c) test was implemented to the public in 1980 for diagnosing diabetes in patients. This test analyzes a percentage of blood sugar attached to hemoglobin for over 3 months (8). This test process is complex, is time-consuming, and requires medical professionals and particular equipment for such to be performed (9).

Type 2 DM (T2DM) is a general diabetes category, and it is categorized as hyperglycemia because of insufficient insulin generation in the human body (10). Recently, numerous diabetes detection approaches have been developed by researchers (11). Many researchers have used datasets that contain lab-test-based medicine indicators for method validation and training process (12), but these prediction approaches considered extensive lab-test-based measurements to diagnose diabetes. However, there is an increasing demand for primary diagnostic solutions that do not depend on measurements of lab test (13). Hence, the research on anthropometric features and its impacts on diabetes prediction approaches has been performed (14). In data-driven diabetes detection solutions, machine learning (ML) has become a popular choice because of its classification capability, which processes statistical techniques without requiring much computation power (15).

The recent developments in diabetes prediction have majorly used ML, an ensemble of learning strategies to enhance the diagnostic accuracy and generalization ability. Simaiya (16) presented the multistage ensemble model integrating classifiers across layers to improve prediction performance on PIMA dataset. Thakur et al. (17) suggested for the model to focus on COVID-19 context, combining advanced feature engineering and ensemble algorithms to enhance prediction accuracy, highlighting the significance of early diagnosis in vulnerable populations. Kaliappan et al. (18) analyzed different datasets by ML algorithms with explainable AI tools such as SHAP and LIME, highlighting the part of model interpretability and data diversity in reliable diabetes prediction. Abousaber et al. (19) addressed the data imbalance issues through integrating ensemble models with oversampling algorithms, obtaining robustness across PIMA and real-world clinical datasets. However, the ML-based algorithms were unable to capture the long-term dependencies in diabetes, and certain research skipped the feature selection process. These drawbacks reduce the performance of diabetes predictions, and to overcome these drawbacks, in this article, the deep learning (DL)-based algorithms were developed, which captures the long-term dependencies effectively. In clinical settings, understanding the features that contribute to the model’s prediction is essential to enable effective decision-making. Hence, this research developed the model for accuracy and also for interpretability in feature level. The significant contributions of the article are described below:

This research paper is organized as follows: Section 2 analyzes the different ML and DL approaches. Section 3 provides the process of diabetes prediction with detailed explanation. Section 4 analyzes the performance of the proposed method and presents results with comparison, and Section 5 concludes the article.

This section analyzed the different ML- and DL-based approaches for diabetes prediction on PIMA, LMCH, and other standard and collected datasets. These approaches are described with their process, advantages, and drawbacks.

Saeed (20) suggested various ML classifiers like gradient boosting, AdaBoost, decision tree (DT), and Extra Tree classifiers for detecting chronic diabetes disease. These methods analyzed PIMA Indian Diabetes dataset (PIMA) and Behavioural Risk Factor Surveillance System (BRFSS) datasets for classifying the patients with positive or negative diagnosis. The suggested ML algorithms effectively predict diabetes, but the ML algorithms cannot capture the long-term dependencies in datasets like glucose levels.

Olisah et al. (21) presented the framework including Spearman correlation and polynomial regression to select features and impute missing values. Various supervised ML algorithms like random forest (RF), support vector machine (SVM), and twice growth deep neural network (2GDNN) were used for classification. The methods were optimized using hyperparameter tuning through grid search and k-fold validation, which were analyzed for its effectiveness to address the prediction issue, but the method does not scale the value in uniform range that reduces the classification performance.

Reza et al. (22) introduced the improved non-linear kernel for the SVM method for enhancing the type 2 diabetes classification. The new kernel utilized the radial basis function (RBF) and RBF city block kernels which enabled the SVM to learn difficult decision boundaries and adapted intricacies. For addressing missing values and outliers, imputation was performed using the median, which ensures the integrity of the dataset. The class imbalance problem was mitigated by leveraging the robust synthetic-based over-sampling technique. However, the SVM approach was unable to capture the long-term dependencies in data, which was crucial for diabetes prediction.

Patro et al. (23) developed the data modeling which depended on correlation measures among features and was utilized for processing the data efficiently to predict diabetes. The standard available Pima Indians Medical Diabetes (PIMA) dataset was used to verify the effectiveness of the developed methods. The developed method predicted diabetes in the early phase and improved the accuracy, but the method does not address the issue of exploding gradients during the training phase, which affects the classification performance.

Dharmarathne et al. (24) implemented the ML interpretation technique called Shapley Additive Explanations (SHAP). Every method exhibited commendable accuracy in detecting patients with diabetes, with the XGB method showing a little edge. By using SHAP, research on the XGB method provided in-depth insights into the reason behind its prediction at a granular phase. The XGB method and local explanation of SHAP were combined to interface the predictions of diabetes in patients. The implemented method reduced the risks associated with diabetes by enhancing awareness, but the method did not consider the feature selection phase, so the whole feature subsets were fed to the classifier, which minimizes the classification performance.

Ejiyi et al. (25) suggested data augmentation and imputing the missing values as the preliminary phases. This method utilized the SHAP for extracting the feature significance and many significant features to fit the Extra Tree (ET), RF, Adaboost, and Xgboost techniques. The SHAP shows that glucose has a specific feature which contributes to many diabetes predictions when integrated with body mass index (BMI) and age. The suggested method effectively predicted diabetes with high performance, but it had less classification performance due to the values in the dataset that were not scaled uniformly.

Bhaskar et al. (26) developed the deep hybrid correlational neural network (CORNN) for detecting diabetes in patients. A few modifications were made to the network layout to enhance the classification accuracy of learning methods. The developed method has enhanced accuracy when compared with non-invasive methods. The CORNN method efficiently classified diabetes with high classification accuracy, but the developed method suffered from the issue of exploding gradients, which leads to less classification performance.

Ganie et al. (27) introduced the five boosting techniques for predicting diabetes on the PIMA diabetes dataset. The dataset was acquired from the University of California Irvine (UCI) ML repository and included numerous significant clinical features. Experimental analysis of data was utilized to identify the data characteristics. Additionally, upsampling, normalization, feature selection, and tuning of hyperparameters were assigned for predictive analysis. The introduced method effectively handled the features in the dataset and improved the performance, but the method does not impute the missing values in the dataset, which leads to less classification performance.

Alnowaiser (28) developed the automatic technique to predict diabetes with concentration on appropriately dealing with missing data and enhancing accuracy. The developed method utilized the K-nearest neighbor (KNN) imputed features with the Tri-ensemble voting classifier method. The developed ensemble method effectively handled the missing values and improved the accuracy but used ML algorithms that cannot capture the long-term dependencies for diabetes prediction.

Tasin et al. (29) implemented the semi-supervised method with XGB used for predicting the insulin features of the standard dataset. The ADASYN and SMOTE techniques are assigned for handling the issues of class imbalance. The ML classification techniques such as DT, SVM, RF, logistic regression (LR), KNN, and different ensemble methods for determining the technique provided good prediction outcomes. However, the method showed less classification performance due to classification errors in the performance.

The existing techniques have some drawbacks like the values in the dataset do not scale uniformly. Some methods do not impute the missing values and then skip the feature selection process. ML approaches were unable to capture the long-term dependencies and did not address the issue of exploding gradients. To overcome these drawbacks from the existing techniques, this article used the min–max normalization technique to scale the values uniformly, and the polynomial regression (PR) technique was used to impute the missing values in the dataset. Then, the RSFS–MPA-based feature selection algorithm is developed to select the relevant features, and then the classification is performed by using Skip-GRU with GC approach. This approach captures the long-term dependencies and mitigates the issue of exploding gradients with the help of the GC method during the training process. These processes help the model to predict diabetes effectively with high performance and accuracy.

The efficient DL-based algorithm is developed in this article to predict diabetes effectively. The datasets used for the diabetes prediction are PIMA and LMCH. Then, the values in the dataset are pre-processed by using the min–max normalization that scales the data into (0,1) range, and the polynomial regression (PR) technique is used for imputing the missing values in data. Next, by developing the RSFS–MPA, the appropriate features are selected and then these are fed into the classification process. For classification, the Skip-GRU with GC approach is developed, which captures the long-term dependencies and effectively predicts diabetes. Figure 1 describes the process of diabetes prediction.

The proposed Skip-GRU with GC approach is simulated with a Python 3.8 environment, and the required system configurations are 8 GB RAM, Windows 10 (64 bit), and i5 processor. The metrics taken to evaluate the performance of Skip-GRU with the GC approach are accuracy, sensitivity, precision, F1-score, and specificity on PIMA and LMCH datasets. The mathematical formula for metrics is given in Equations 27-31.

Accuracy—This is the ratio of the whole number of accurate predictions to the whole number of predictions, and its mathematical formula is given in Equation 21.

Sensitivity—This is the ratio of patients with diabetes (positive instances) who are accurately identified as diabetic, and it is calculated as a proportion of true positives (TP) to the sum of TP and false negatives (FN). The mathematical formula for sensitivity is given as Equation 28:

Specificity—This is the ratio of patients without diabetes (negative instances) who are accurately identified as non-diabetic, and it is calculated as the proportion of true negatives (TN) to the sum of TN and false positives (FP). The mathematical formula for specificity is given as Equation 29:

Precision—This is the ratio of patients with diabetes, positive instances, who are accurately identified as diabetic from whole diabetic patients, and it is executed as a proportion of TP to the sum of TP and FP. The mathematical formula for precision is given as Equation 30:

F1-score—This is the average of precision and sensitivity, and it considers the value of both FP and FN. It is given as Equation 31:

Table 2 presents the explored ranges and final selected values for every hyperparameter. These values are chosen based on empirical testing using the PIMA and LMCH datasets. A smaller learning rate of 0.001 stabilized the training and improved the convergence. It ensures smooth and well-controlled weight updates when larger learning rates resulted in oscillations and suboptimal convergence. A hidden unit size of 128 provides better trade-off and offers effective learning ability and improved generalization ability. Fewer hidden units lead to underfitting, while a larger number of hidden units causes overfitting. A dropout rate of 0.3 prevents overfitting by randomly deactivating neurons in training, thereby enhancing generalization. The higher dropout rates introduced excessive regularization, which degrades performance through limiting the model’s ability to learn complex feature interactions. A weight decay value of 0.1 provides strong regularization that penalized large weight magnitudes and effectively constrains model complexity, thereby reducing overfitting. Lower weight decay values provide ineffective regularization, while larger values cause underfitting through the excessive restriction of the model’s learning ability. A batch size of 32 offers a better trade-off between stable gradient estimates and memory efficacy. Smaller batch sizes introduce a high variance in gradient updates, causing noisy and unstable training, while larger batches reduce gradient noise, but slows down convergence and increases the generalization error because of less weight updates. A gradient clipping threshold of 1.0 effectively mitigates the exploding gradient problem in the training process of the Skip-GRU model. It constrains the norm of gradients in a controlled range and provides stable and consistent parameter updates. Lower threshold values reduce the gradient flow, leading to slow learning and poor convergence, while higher values fail to prevent large updates, resulting in training instability.

In Table 3 , the performance of the feature selection algorithm is evaluated on the PIMA and LMCH datasets with different metrics. The different feature selection algorithms considered for evaluating the developed RSFS-MPA are the reptile search algorithm (RSA), crow search algorithm (CSO), crayfish optimization algorithm (COA), and traditional MPA. The RSFS is incorporated with MPA for the feature selection process, which improves the search ability and effectively balances the local and global processes of traditional MPA. By using this strategy, the MPA effectively searches for the best feature subset and chooses the best features for classification. The developed RSFS-MPA achieved 97.84% sensitivity, 97.35% precision, 97.59% F1-score, 98.23% accuracy, and 97.61% specificity on the PIMA dataset and 97.21% sensitivity, 97.46% precision, 97.33% F1-score, 97.65% accuracy, and 97.53% specificity on the LMCH dataset.

In Table 4 , the performance of the classifier is evaluated on the PIMA and LMCH datasets with actual features based on different metrics. The different classifiers considered for evaluating the Skip-GRU with GC approach are recurrent neural network (RNN), long short-term memory (LSTM), GRU, and Skip-GRU. The proposed Skip-GRU with GC achieved 96.25% sensitivity, 95.87% precision, 96.05% F1-score, 96.71% accuracy, and 95.46% specificity on the PIMA dataset and 95.19% sensitivity, 95.32% precision, 95.25% F1-score, 95.25% accuracy, and 95.43% specificity on the LMCH dataset.

In Table 5 , the performance of the classifier is evaluated on the PIMA and LMCH datasets after feature selection based on different metrics. The different classifiers considered for evaluating the Skip-GRU with GC approach are RNN, LSTM, GRU, and Skip-GRU. Here the relevant features from the whole feature subset are selected by using the RSFS-MPA feature selection technique. By eliminating the irrelevant features, only the relevant features are fed into the classification process and help to enhance the classification performance. By using the Skip-GRU approach, unwanted features were skipped and only the significant features for classification were fed. Then, the GC technique is used during the training process of Skip-GRU, which helps to mitigate the issue of exploding gradients and enhances the process of Skip-GRU for diabetes prediction. The RFFS-MPA effectively filters out inappropriate and redundant features by minimizing noise. This selected feature refinement resulted in a much focused training process and minimized the overfitting issue, which is especially significant in medical datasets that contain correlated and low-variance features. The Skip-GRU model enhances the performance through dynamically ignoring irrelevant inputs by the jump probability mechanism. This enables a model to adaptively control data flow and allows essential temporal inputs to influence the hidden states. From the results, Skip-GRU consistently outperformed existing models like RNN, LSTM, and GRU across the entire evaluation metrics. The incorporation of GC to Skip-GRU model provides stability in the training process through gradient magnitudes, which efficiently addresses the exploding gradient issue. This is particularly essential while dealing with high-dimensional data, as the uncontrolled gradients cause unstable updates and degrading performance. The proposed Skip-GRU with GC approach achieved 97.84% sensitivity, 97.35% precision, 97.59% F1-score, 98.23% accuracy, and 97.61% specificity on the PIMA dataset and 97.21% sensitivity, 97.46% precision, 97.33% F1-score, 97.65% accuracy, and 97.53% specificity on the LMCH dataset.

In Table 6 , the performance of Skip-GRU with the GC approach is evaluated based on k-fold validation. It splits the dataset into multiple subsets and uses every fold as a training and validation set in various iterations. Instead of training and testing on similar data, the k-fold cross-validation gives a much more reliable estimation of method performance on unseen data. The performance of Skip-GRU with GC approach is evaluated for k-values of 2, 3, 4, 5, and 6. In that, K = 5 has achieved the highest values when compared to other k-fold values and provides better balance among training and validation coverage. This analysis is essential to validate the model’s robustness, stability, and generalization ability over various training–testing splits. In K-fold cross-validation, the dataset is divided into K folds and the model is trained K times, every time using K-1 folds for training and remaining fold for testing. This ensures that each instance in the dataset is utilized for training and validation and outcomes are averaged to overcome variance because of random data splits. The results show that Skip-GRU with GC model provides a consistent performance over different K-fold configurations. The less fluctuation across folds determines that the model is not sensitive to data partitioning.

This research developed an effective DL-based approach for enhancing diabetes prediction performance using the PIMA and LMCH datasets. The processes involved in diabetes prediction are pre-processing the data using min–max normalization and PR and then the feature selection using the RSFS-MPA. Finally, the data is classified by using the Skip-GRU with GC approach. The RSFS used in traditional MPA enhanced the performance of traditional MPA for feature selection. This process eliminates the irrelevant features from the whole feature subset and selects only the relevant features and feeds them to the classification process. The Skip-GRU approach effectively classifies diabetes with high accuracy. The GC technique is introduced in the Skip-GRU approach during the training phase, which mitigates the exploding gradient issue and enhances the performance of diabetes prediction. The proposed Skip-GRU with GC approach achieved 98.23% accuracy on the IMA dataset and 97.65% accuracy on the LMCH dataset when compared to existing approaches like the Extra Tree classifier and seven-layer-deep CNN.