Introduction

The World Health Organization has reported that approximately 1.6 million deaths occur annually due to diabetes1. Diabetes, characterized by insufficient insulin, is a complex condition affecting various organs such as the liver, muscles, and adipose tissue, as insulin regulates glucose production2. Inadequate insulin levels result in elevated blood glucose, leading to potential organ damage, including the eyes, cardiovascular system, and nerves3. Diabetes manifests as various types, primarily Type 1, Type 2, and Gestational4. Type 2 diabetes often stems from obesity-induced insulin resistance, while type 1 diabetes arises from immune system attacks on pancreatic beta cells5. Both types of hyperglycemia are due to insulin deficiency or resistance6. Type 1 diabetes predominantly affects children and young adults, whereas type 2 diabetes is prevalent in adults7. Gestational diabetes, diagnosed during pregnancy, poses risks to both maternal and fetal health but can be mitigated with early detection and management8. Monitoring blood glucose levels and adhering to medical advice are crucial for diabetic individuals9. The typical blood glucose level can vary based on factors such as age, family medical history, individual health status, and dietary habits. Glucose, derived from food, serves as the primary energy source for the body. Various laboratory examinations, including tests for HbA1c, C-peptide, glucose, insulin, acetone, and blood gases, are necessary for diabetes detection. A positive result in the urine indicated the presence of acetone. Hemoglobin A1c (HbA1c) serves as a diagnostic tool, reflecting the average glucose level over the past two to three months10. The C-peptide test aids in identifying pancreatic cell destruction by the immune system, particularly in pediatric diabetes patients. Abnormal C-peptide levels may indicate poor insulin utilization by the body. Blood gas analysis accurately measures oxygen and carbon dioxide levels within the body. Diabetes manifests as several chronic symptoms such as excessive hunger, thirst, frequent urination, neuropathy, nephropathy, retinopathy, cardiovascular issues, peripheral vascular disease, dental complications, stroke, diabetic foot syndrome, encephalopathy, hyperthyroidism, adrenal gland tumors, cirrhosis of the liver, glucagonoma, persistent vomiting, severe abdominal pain, metabolic acidosis, fatigue, dysuria, diabetic ketoacidosis, weight loss, and various other health complications. Early detection facilitates effective diabetes management, involving a balanced fitness regimen, proper treatment, and nutritious eating habits11. Preventive measures encompass addressing risk factors and fostering healthy lifestyles, including regular physical activity, limited intake of processed foods and refined carbohydrates, adequate protein consumption, and ample fruits and vegetables. Risk factors include maternal glucose levels during pregnancy and childhood obesity. Type 1 diabetes treatment typically involves insulin administration via injections or insulin pumps, frequent blood glucose monitoring, and carbohydrate counting. Additionally, hyperbaric oxygen therapy has recently gained attention for its ability to enhance healing processes by elevating oxygen levels in the body. Pressurized oxygen serves as a complementary therapeutic approach for diabetic foot care, facilitating increased blood flow and oxygen delivery to damaged cells and tissues, thereby aiding in their restoration and effective recovery. Recently, the proliferation of medical knowledge has surged alongside advancements in artificial intelligence methodologies, which are adept at managing and processing this wealth of information to effectively enhance diabetes management practices. The fundamental objective of intelligent learning from accumulated data is to enable computers to autonomously acquire knowledge without human intervention. ML has emerged as a computational technique that harnesses experiences, historical data, and repository information to cultivate intelligence12. ML proves invaluable in diagnosing diabetes; it boasts various types of learning, such as supervised learning, unsupervised learning, and reinforcement learning, as depicted in Fig. 1. Supervised learning algorithms are trained using examples, incorporating both input and target data into the training dataset. Classification and regression exemplify supervised learning techniques. Conversely, unsupervised learning algorithms discern latent patterns within training data without relying on labeled information to analyze and extract significant features. Clustering and dimensionality reduction represent unsupervised learning methodologies. Reinforcement learning empowers agent systems to learn within interactive environments, guided by feedback mechanisms employing reward and penalty policies. Rewards, whether positive or negative, shape the behavior of the agent or system, with applications ranging from robot navigation to real-time decision-making13.

Fig. 1
Fig. 1
Full size image

Machine learning types.

The remainder of this paper is organized as follows: in the second section, we present the related work. Section Methods details the proposed architecture and the corresponding features through a comparison of nine classification techniques. The results are presented in Section Train phase, and the conclusions are presented in Sect.on Machine learning multi-class classification.

Related work

This study introduces two ensemble models based on stacking techniques, one employing traditional ML algorithms and the other utilizing a deep neural network (DNN) for the early detection of diabetes14. The models are developed by integrating three data sources: the Pima Indian Diabetes (PID) dataset, a simulated dataset, and a local healthcare dataset. By combining the predictions of multiple base classifiers, the ensemble models aim to improve predictive accuracy and robustness. Performance evaluation was conducted using both a train-test split and 5 fold cross-validation. The DNN-based ensemble demonstrated the highest performance on the simulated dataset, achieving an accuracy of 95.5%, precision of 94%, recall of 97%, and an F1-score of 96%. On the PID dataset, the stacked ML models achieved an accuracy of 75.03% with the train-test split and 77.10% with cross-validation. Overall, the proposed approach achieved accuracy levels between 92 and 95%, with corresponding precision, recall, and F1-scores ranging from 88 to 96%, highlighting its potential for reliable early-stage diabetes detection.s

This study seeks to improve the generalizability and predictive accuracy of diabetes classification models by addressing the limitations inherent in single-population datasets15. To this end, the researchers integrated two publicly available datasets: the Pima Indian Diabetes (PID) dataset comprising 768 samples, and the German Society dataset containing 2000 samples. Ensemble learning methods, specifically RF and GB, were employed to evaluate model performance on both individual and combined datasets. The findings indicated that models trained on the merged dataset outperformed those trained on individual datasets, with GB and RF achieving accuracies of 0.991 and 0.988, respectively, compared to RF of 0.817 on PID and GB of 0.996 on the German dataset.

This study proposes a pipeline-based multi-classification framework for predicting diabetes status categorized as diabetic, non-diabetic, and prediabetic using an imbalanced dataset from Iraqi patients16. To overcome challenges such as limited labelled data, missing values, and class imbalance, the framework incorporates extensive preprocessing steps, including duplicate removal, imputation, normalization, standardization, feature selection, and k-fold cross-validation. Several ML algorithms KNN, SVM, DT, RF, AB, and NB are applied, with a weighted ensemble model guided by AUC introduced to enhance predictive performance. Hyperparameter optimization is conducted via grid search and Bayesian methods. The proposed ensemble model outperforms individual classifiers, achieving an accuracy of 0.9887, a precision of 0.9861, a recall of 0.9792, an F1-score of 0.9851, and an AUC of 0.999. These findings indicate the framework’s robustness and scalability, with strong potential for broader application in diabetes prediction across diverse populations.

This study presents an automated diabetes prediction system leveraging ML techniques on a combined dataset comprising the PID dataset and additional records from 203 female patients in Bangladesh17. Mutual information is utilized for feature selection, and a semi-supervised learning approach employing Extreme Gradient Boosting (XGBoost) is adopted to estimate insulin levels. To mitigate class imbalance, oversampling techniques such as SMOTE and ADASYN are applied. A range of classifiers, including DT, SVM, RF, LR, KNN, and several ensemble models, are evaluated. The XGBoost model combined with ADASYN achieves the highest performance, with 81% accuracy, an F1-score of 0.81, and an AUC of 0.84. To assess generalizability, domain adaptation techniques are employed, while model interpretability is addressed using explainable tools such as SHAP and LIME. The system is further integrated into both a web-based interface and an Android application, enabling real-time diabetes prediction and enhancing its applicability in clinical and remote healthcare settings.

This study focuses on early diabetes prediction through the application of nature-inspired metaheuristic algorithms and includes both case studies and a comprehensive review of existing predictive models18. Various bio-inspired optimization techniques, such as Ant Colony Optimization, Bat Algorithm, Cuttlefish Algorithm, Elephant Herd Optimization, and Artificial Bee Colony, are employed to enhance classifier performance via hyperparameter tuning. A hybrid Bat Algorithm is specifically used to optimize several classification models, while SMOTE is applied to address class imbalance. Among the evaluated models, a voting classifier enhanced with both SMOTE and the Bat Algorithm achieved the highest accuracy of 98%. The findings highlight the effectiveness of nature-inspired optimization strategies in improving prediction performance and further explore the integration of dietary recommendations based on predictive outcomes, suggesting a promising direction for personalized diabetes management.

This study emphasizes the early prediction of diabetes using both statistical and non-statistical ML techniques, applied to the PID Dataset comprising 768 patient records19. Key predictive features include age, BMI, and blood glucose levels, which are recognized as critical risk indicators. A diverse set of classification algorithms, including LR, DT, RF, KNN, NB, SVM, GB, and ANN, is evaluated for its predictive effectiveness. Among these, the ANN model demonstrated the highest accuracy at 78.57%, followed by RF at 76.30%. The results highlight the potential of ML as a robust, data-driven approach for early diabetes risk assessment, enabling timely medical intervention and improved disease management.

This study introduces an advanced diabetes prediction framework that integrates conventional ML models, including LR, SVM, NB, and RF, with ensemble methods such as AB, GB, Extra Trees, and XGBoost20. Central to the framework is DNet, a novel hybrid deep learning architecture that combines Convolutional Neural Networks (CNN) for feature extraction with Long Short-Term Memory (LSTM) layers to capture temporal dependencies. DNet incorporates convolutional and residual blocks with skip connections, as well as Batch Normalization and Dropout for improved regularization and generalization. The framework is evaluated on a real-world Kaggle diabetes dataset using cross-validation and performance metrics such as precision, recall, F1-score, and ROC-AUC. DNet significantly outperforms all baseline models, achieving an accuracy of 99.79% and an AUC-ROC of 99.98%. These results underscore the model’s robustness and its strong potential for real-world application in early and accurate diabetes diagnosis.

ANN was utilized for diabetes prediction, with model training, validation, and testing conducted using the Just Neural Network software environment21. The dataset consisted of 1004 records with 9 input features, collected from the Association of Diabetic City of Urmia. Training was performed using the backpropagation algorithm to reduce prediction error. The model achieved an average error rate of 0.01 and a prediction accuracy of 85.09%, demonstrating its effectiveness in identifying diabetic cases.

The dataset attributes, including age, gender, family history, BMI, and blood glucose levels, exhibit interrelated patterns relevant to diabetes prediction22. ML algorithms were applied for the detection and classification of diabetes, with further potential for enhancing predictive accuracy through the identification of diabetes type and the estimation of comorbid disease risk. Two datasets, the PIMA dataset and a Clinical Survey (CS) dataset, were utilized in the study. Diabetes classification into pre-diabetes and diabetes categories was performed using multiple classification algorithms. Preprocessing methods such as data augmentation and sampling were employed to improve model performance. Among the tested algorithms, including RF, Light Gradient Boosting Machine (LGBM), GB, SVM, DT, and XGBoost, the LGBM classifier achieved the highest accuracy of 95.20%.

Researchers employed various machine learning models including ANN, KNN, NB, AB, LR, RF based on DT, and SVM to predict diabetes23. Using Pearson’s correlation for feature analysis and the WEKA tool for data processing, missing values in the PID dataset were imputed with mean values. The study achieved an accuracy of 83% with a low error rate.

Using the PID dataset, researchers applied ANN, data mining, and ML algorithms to predict diabetes24. Missing values were imputed with mean values, LR and SVM models achieved accuracies of 78.86% and 78.29%, respectively. The ANN model, configured with multiple epochs and two hidden layers utilizing the ReLU activation function, demonstrated superior performance with an accuracy of 90.1%.

Building on prior research, it is noted that all patients in the widely utilized PIMA dataset are female, have experienced pregnancies, and are aged 21 years or older. Due to the prevalence of missing values in this dataset, thorough preprocessing is essential to prevent inaccurate predictions. This study focuses on the early detection of pediatric diabetes and aims to identify the key factors affecting the health of a younger patient population. Accordingly, the researchers seek to develop an optimized machine learning algorithm designed to predict diabetes with improved accuracy.

Methods

This study introduces a novel model for multiclassification of diabetes types. The model encompasses a series of steps, beginning with data collection and proceeding through preprocessing, and data analysis-based feature extraction followed by a training/testing phase. The next phase of the ML multiclassification process is followed by evaluating the classifier’s performance, optimization, and prediction phases. The system architecture is graphically presented in Fig. 2.

Fig. 2
Fig. 2
Full size image

The diabetes data types system architecture.

Data collection

Data forms the foundation of ML models, shaping their effectiveness. Gathering the appropriate data, both in terms of quality and quantity, is pivotal for constructing an optimal model. The reliability of the collected data plays a crucial role across all phases of model classification, ultimately influencing the quality of predictions. This phase involved selecting significant data features and determining the necessary sample size, guided by assumptions regarding the most relevant attributes to diabetes. Therefore, the quality of data directly impacts the performance of the model, contributing to its overall efficacy. We declare the detailed explanation of how to unify different datasets into the new DTD dataset as shown in online supplementary material.

Dataset

The proposed system employs the newly developed DTD dataset, which integrates data from four distinct sources: the Pediatrics, PID, Pone, and Gestational diabetes datasets, resulting in a comprehensive dataset of 5312 patients with 13 attributes including Age, Sex, BPressure, NPregnancies, BMI, HbA1c, Insulin, POGTT, FOGTT, PGlucose, FGlucose, Diagnosis and DiagnosisType. The Pediatrics dataset, sourced from Mansoura University Children’s Hospital in Egypt, includes 619 patients aged 1 to 19. The PID dataset, obtained from the UCI Repository, contributes data from 768 patients. The Pone dataset, collected from hospitals in Thanjavur district, Tamil Nadu, India, includes 400 patients with 23 features. The Gestational dataset comprises data from 3012 patients with 17 attributes, gathered by diabetes researchers. The goal of the DTD dataset is to identify key factors influencing diabetes occurrence within a multiclass classification framework. We use two external datasets: the first external dataset is diabetes_prediction_dataset, containing 100,000 patient records and 9 features, namely gender, age, hypertension, heart_disease, smoking_history, bmi, HbA1c_level, blood_glucose_level, and diabetes which was also utilized, comprising 91,501 non-diabetic and 8499 diabetic cases. The second external dataset is diabetes_Dataset, containing 34 features namely: Target, Genetic Markers, Autoantibodies, Family History, Environmental Factors, Insulin Levels, Age, BMI, Physical Activity, Dietary Habits, Blood Pressure, Cholesterol Levels, Waist Circumference, Blood Glucose Levels, Ethnicity, Socioeco2mic Factors, Smoking Status, Alcohol Consumption, Glucose Tolerance Test, History of PCOS, Previous Gestational Diabetes, Pregnancy History, Weight Gain During Pregnancy, Pancreatic Health, Pulmonary Function, Cystic Fibrosis Diagnosis, Steroid Use History, Genetic Testing, Neurological Assessments, Liver Function Tests, Digestive Enzyme Levels, Urine Test, Birth Weight, and Early Onset Symptoms used to predict 12 distinct diabetes types and prediabetic. The diabetes types are namely: Type 1, Type 2, Type 3c, Gestational, MODY, LADA, Secondary, Neonatal Mellitus, Wolcott-Rallison Syndrome, Steroid-Induced, Cystic Fibrosis-Related, and Wolfram Syndrome.

Figure 3 shows the binary distribution of diabetes (2317 diabetic and 2995 non-diabetic patients), while Fig. 4 outlines the composition of the four source datasets. Additionally, a rich Kaggle diabetes dataset covering various forms of the disease—such as Steroid-Induced, Neonatal, Prediabetes, Type 1, and Wolfram Syndrome—was referenced to support broader analysis of genetic, lifestyle, and medical factors contributing to diabetes.

Fig. 3
Fig. 3
Full size image

The number of diabetes and non-diabetes patients.

Fig. 4
Fig. 4
Full size image

The PIMA, pediatrics, CS, and gestational diabetes datasets.

Table 1 presents the DTD dataset, which comprises thirteen attributes and their comprehensive descriptions. The diagnosis attribute serves as the dependent output variable, while the remaining thirteen attributes are considered independent input features. The DiagnosisType label is utilized for multiclassification purposes, as depicted in Fig. 5.

Table 1 The attribute of the DTD dataset.
Full size table
Fig. 5
Fig. 5
Full size image

The diagnosis types.

Data preprocessing

Some errors can result from human mistakes during the previous data collection phase. This led us to perform a preprocessing step that included adjusting the data format, handling missing values, feature selection, data sampling, and feature scaling.

Data format

The current phase is concerned with manipulating the collected input attributes to be in a clear and correct format. The preprocessing phase is used to organize and clean the data for further analysis and processing. This approach assists in the accuracy and precision of interpreting data features via the ML classifier algorithms. We prepared the DTD dataset in CSV file format.

Missing values

An empty value within the attributes of the DTD dataset indicates a missing value, which is typically denoted by null indicators. Missing data samples may arise from errors during the data collection phase or from unperformed analysis requests. Such missing values can detrimentally impact the overall performance of the system25. In patient records, missing values may occur for one or multiple attributes for a defined percentage of patients. Addressing the issue of missing values can be approached in two ways. First, one may opt to eliminate features with missing values, although this risks discarding pertinent information and reducing the dataset size. Second, one can replace missing values by Multiple Imputation by Chained Equations (MICE) for the DTD dataset26. MICE is widely used to handle missing data in datasets that contain both numerical (continuous) and categorical (discrete) variables. The imputation used in MICE is a regression-based model where each missing value Yj is predicted using the observed values of other variables27. The general form for imputation is indicated using Eq. (1):

$$Y_{j}^{m} = f\left( {X^{m} ,\theta_{j} } \right)$$
(1)

where \({\text{Yj}}^{m}\) is the imputed value of variable j in the m-th imputed dataset, Xm represents the observed values of all other variables in the dataset for the m-th imputed dataset. f is a regression model (e.g., linear regression or LR) that predicts \({\text{Yj}}^{m}\) based on Xm. The structure of the DTD dataset multiclass classification includes numerical features like Age, BMI, BPressure, NPregnancies, HbA1c, FGlucose, PGlucose, FOGTT, POGTT, Insulin, and categorical features (e.g., Sex, Diagnosis). The target classes (e.g., Normal, Type 1, Type 2, or gestational diabetes). A sample of the DTD dataset is depicted in Fig. 6. where missing values are indicated by white cells. The number of missing values in the DTD dataset is detailed in Table 2. Then we standardize the data using Standard Scaler.

Fig. 6
Fig. 6
Full size image

The dataset mini-batches.

Table 2 The missing values of The DTD dataset.
Full size table

Handle class imbalance

We use the Synthetic Minority Over-sampling Technique) SMOTE) A technique to balance imbalanced datasets by creating synthetic samples of the minority class35. Instead of duplicating rows, it generates new data points using interpolation. It’s used before training classifiers to prevent the model from being biased toward the majority classes. SMOTE helps by equalizing the number of samples per class and improving model generalization across all classes. The DTD samples are distributed as follows:—Normal patients are: 3003, Type 1 patients are: 277, Type 2 patients are: 659, and Gestational patients are: 1373. Using the SMOTE technique, we balance the dataset to be: Normal patients are 3004, Type 1 patients are 3004, Type 2 patients are 3004, and Gestational patients are 3004, to prevent overfitting.

Feature selection

Feature selection plays a crucial role in the feature extraction process, aiming to eliminate redundant features and retain those essential for constructing an efficient predictive model, thereby enhancing classification accuracy. This process aids in comprehending the significance of all extracted features, allowing for the utilization of valuable values while discarding outlier features. By prioritizing features with the highest correlation and importance scores, the feature selection method minimizes the execution time and mitigates the risk of data overfitting. Identifying interactions among input variables that influence system output performance is of paramount importance.

To analyze correlations between different attribute types (numerical and categorical), we apply two statistical techniques One-Way ANOVA test and the Chi-square tests. The ANOVA test is used to compare the means of a numerical variable across multiple categories of DiagnosisType (target classes). ANOVA checks if the mean Age, BMI, BPressure, NPregnancies, HbA1c, FGlucose, PGlucose, FOGTT, POGTT or Insulin differs significantly across DiagnosisType. It helps to assess whether the variation in the data is due to differences between groups or if it’s just random variation. To perform ANOVA test, state the null hypothesis (H₀): The means of the groups are equal, using Eq. (2):

$${\text{H}}_{0} : \, \mu_{{1}} = \, \mu_{{2}} = \, \mu_{{3}} = \, \mu_{{4}} = \ldots \ldots . \, \mu_{{\text{k}}}$$
(2)

where k is the total number of groups.

Alternative hypothesis (H₁) at least one of the group means is different from the others using Eq. (3):

$${\text{H}}_{1} {:}\mu_{{\text{i}}}$$
(3)

where I ∈ {1, 2, 3, …, k}.

The F-statistic (F) is calculated by comparing the variance between groups and within groups using Eq. (4):

$${\mathbf{F}} = \frac{{{\mathbf{Variance}} {\mathbf{between}} {\mathbf{Groups}} }}{{{\mathbf{Variance}} {\mathbf{within}} {\mathbf{Groups}}}}$$
(4)

We use the p-value statistical measure that helps to determine the significance of your results in a hypothesis test. It tells the probability of obtaining results at least as extreme as the observed results, if the null hypothesis (H0) is true. All the p-values are very small (< 0.0001), with many being close to zero, indicating that there is a statistically significant difference in Age, BPressure, NPregnancies, BMI, HbA1c, POGTT, FOGTT, PGlucose, FGlucose, and Insulin across DiagnosisType as indicated in Table 3.

Table 3 F-statistic and p value of numerical attributes.
Full size table

We also apply the Chi-Square Test (χ2) to evaluate whether two categorical variables are statistically associated. This test is especially suitable when you want to examine if DiagnosisType is related to categorical features like Sex or Diagnosis. In practice, the test compares each observed cell frequency with its expected count under the null hypothesis of independence. The Formula for the Chi-Square Test using Eq. (5):

$$X^{2} = \sum \frac{{\left( {O - E} \right)^{2} }}{E}$$
(5)

where O is the actual count in each category, E is the expected count if there is no association.

The degrees of freedom (D) using Eq. (6):

$${\text{D }} = \, \left( {{\text{R}} - {1}} \right){\text{ x }}\left( {{\text{C}} - {1}} \right)$$
(6)

where R is the number of rows and C is the number of columns.

The Chi-square statistic is notably large, and the p-value is less than 0.0001 well below the conventional threshold of 0.05, indicating a significant association between the Sex and DiagnosisType variables. Similarly, the Chi-square between Diagnosis and DiagnosisType is also very large, with a p-value less than 0.0001, confirming a significant association, as shown in Table 4.

Table 4 Chi-square, P value, and degrees of freedom of sex and diagnosis with diagnosistype.
Full size table

Data sampling

We partition the training dataset into smaller chunks. This chunk of the DTD dataset is divided into smaller batches aids in the training phase, as illustrated in Fig. 6.

Train phase

This phase focuses on training multiple classifiers using a set of input attributes prepared during the data preprocessing stage. Classifiers such as ANN, LR, NB, DT, AB, RF, GB, KNN, and ANN are each individually trained on the dataset. The objective is to build models capable of accurately classifying unseen data into one of four categories: Type 1 Diabetes, Type 2 Diabetes, Gestational Diabetes, or Normal based on the predefined diagnosisType label. The dataset, containing 5312 records, is split into two primary subsets: 70% (3718 samples) for training and 30% (1,594 samples) for testing, as illustrated in Fig. 7. The training set is further subjected to 5 fold cross-validation to enhance model reliability34. In each fold, approximately 56% of the total dataset is used for training, while about 14% is used for validation. This process is repeated five times, with each fold serving as the validation set once. Stratified sampling is applied during splitting to maintain balanced class distributions. The test set remains untouched throughout the cross-validation process and is used only for final performance evaluation.

Fig. 7
Fig. 7
Full size image

The train, validation and test five k-folds.

Machine learning multi-class classification

The pivotal aspect of the multiclass classification process is training the classifier. This step involves creating a model by training it with nine classifiers, which will subsequently be utilized to classify unlabeled “DiagnosisType” into the four output categories Normal, Type 1, Type 2, Gestational patients. We proceed by training nine ML classifiers, including ANN, LR, NB, DT, AB, RF, GB, ET and KNN34. These supervised ML algorithms are specifically chosen to perform multiclassification on the DTD dataset. Following this, the nine ML classifiers are compared, and the most suitable ones for the dataset are selected. The output of this phase is a trained classifier referred to as a model, which is prepared for testing. The parameters are fine-tuned on the developed predictive model, and performance measures are calculated, resulting in the construction of a superior ML model with a highly accurate performance level.

ANN classification algorithm

Previously, we conducted binary classification to distinguish between diabetic patients and non-diabetic patients. Subsequently, in our proposed model, we utilized nine supervised ML techniques to perform multiclassification of diabetes type prediction using an ANN. An ANN typically comprises multiple layers of units known as neurons. In each mini batch, the input features from the DTD dataset are forwarded to the initial input layer. The neurons in the first layer receive a vector composed of twelve input features, while the subsequent hidden layer’s neurons are linked to the input layer’s neurons through a combination of weights and the Rectified Linear Unit (ReLU) activation function37. The neurons in the last output layer receive a combination of outputs with corresponding weights and apply the SoftMax activation function. The ANN is configured for four output nodes: normal patient, Type 1, Type 2 and gestational diabetes patients as shown in Fig. 8. The cross-entropy of the multinomial distribution serves as the cost function, measuring the disparity between predicted and actual outputs to adjust weights and biases accordingly. We use the swarm optimization for fine-tuning hyperparameters36. The ADAM algorithm is employed to update the assigned weight and bias values, with a regularization factor of 0.01 and 50 epochs, respectively. The number of iterations needed to complete one epoch corresponds to the number of batches. This iterative process continues until the desired output aligns closely with the actual output. The SoftMax function aids in the final classification with multiple output probabilities (0, 1, 2, or 3) for different patients, as depicted in Fig. 10. With this, our training concludes, yielding a multiclassification prediction with high accuracy. The ANN Algorithm is described in online Appendix B.

Fig. 8
Fig. 8
Full size image

ANN architecture.

Fig. 9
Fig. 9
Full size image

The SHAP of DTD dataset.

Fig. 10
Fig. 10
Full size image

Age and diagnosis type attributes.

Test phase

The multiclassification technique was subsequently applied to accurately assign a class label to unlabelled DiagnosisType, effectively distinguishing between Type 1, Type 2, Gestational diabetes, and normal cases. A mapping function is utilized to classify the unlabelled DiagnosisType and ascertain its appropriate label. It is imperative to compute the probability of assigning the patient to the respective class label using eight multiclassification techniques.

Optimization

We utilize the Keras and TensorFlow libraries to construct a model for the ANN. The training–testing split and cross-validation functions from the Sci-Kit-Learn library are employed for data splitting. Various ML algorithms, including LR, NB, DT, AB, RF, GBC, ET, KNN, and ANN, are implemented in this phase. The primary objective is to enhance multiclassification efficiency, accuracy, and reduce computational complexity compared to manual methods. The ANN model is trained with tuned hyperparameters using the Particle Swarm Optimization (PSO) technique to determine the best-fit probable value for model optimization. PSO plays a crucial role in improving the model’s performance by adjusting the particles’ positions in the swarm to find the optimal set of parameters. PSO optimizes several key parameters, such as the number of hidden units, batch size, learning rate, momentum, and the number of hidden layers. In this case, PSO will search for the optimal number of neurons. The batch size, which determines the number of training examples processed in each iteration before updating the model’s weights, will be optimized between 5 and 64. PSO also searches for the best learning rate (lr) within the range of 0.0001 to 0.1, using Adam optimizer to minimize output error during backpropagation, which controls how quickly the model updates its weights during training.

Evaluation

This phase focuses on evaluating the performance of each classifier by comparing predicted class labels with the actual labels using the trained models. To interpret and understand model behaviour, we employ SHapley Additive exPlanations (SHAP) Summary Plots, which illustrate the contribution of each feature to individual predictions as well as the overall model output as shown in Fig. 9. These plots enhance transparency by highlighting which features influence predictions positively or negatively, thereby improving interpretability and trust in the model. Additionally, we apply Particle Swarm Optimization (PSO) for hyperparameter tuning to optimize each classifier’s performance. The effectiveness of the machine learning algorithms is assessed using various evaluation metrics, including precision, Mean Squared Error (MSE), R2 score, training accuracy and AUC3.

Prediction

The performance of all classifiers is compared to determine the most effective model for recommendation. The selected classifier demonstrating the highest predictive accuracy serves as the optimal model, significantly enhancing the system’s overall capability to predict class labels for new data. To support this evaluation, three key visualizations are generated: a Confusion Matrix applied to the full dataset, ROC Curves for each class to assess classification performance, and a Learning Curve that illustrates the relationship between training size and both training and validation accuracy. Finally, the best model is used to predict the DiagnosisType on an external dataset namely “diabetes_prediction_dataset”. The proposed system is designed to use a trained and optimized classifier to make predictions on a new, external dataset that was not part of the original training data. It first loads the external dataset from a CSV file, removes the target column if it exists (since the goal is to predict it), and ensures that the data’s structure matches the training features. It then applies the same preprocessing steps used during training such as imputing missing values and standardizing the data to prepare it for prediction. Using the final model, which was optimized through PSO, it generates predictions for each sample, mapping the predicted numeric labels to meaningful class names like “Normal,” “Type1,” “Type2,” or “Gestational.” Finally, the results, including the original features and predicted labels, are saved to a new CSV file for review or further analysis following the procedure outlined in Algorithm 1.

Algorithm 1
Algorithm 1
Full size image

Predict and Interpret Using Trained ML Model

Results and discussion

Early detection of diabetes significantly improves clinical outcomes and facilitates effective long-term disease management. In this study, a supervised ML system was developed to classify multiple types of diabetes. The proposed ML model is designed for seamless integration into routine clinical workflows and self-management protocols, supporting both diagnosis and risk assessment of various diabetes types. The model was trained and evaluated on the DTD dataset, a composite dataset that integrates patient records from four distinct sources: Pediatrics, PID, Pone, and Gestational Diabetes datasets. This heterogeneous dataset enabled a multiclass classification task covering four categories: Type 1 Diabetes, Type 2 Diabetes, Gestational Diabetes, and Normal (non-diabetic). A comprehensive preprocessing phase was conducted to prepare the dataset for modelling. This included standardization of file formats, removal of outliers and inconsistencies, and imputation of missing values using the MICE (Multivariate Imputation by Chained Equations) technique, applied to both numerical and categorical features. To address class imbalance within the dataset, the SMOTE (Synthetic Minority Over-sampling Technique) method was used, generating synthetic samples for underrepresented classes and improving classifier fairness. Nine supervised ML algorithms were employed for classification: LR, NB, DT, RF, AB, GB, SVM, KNN, and ANN. To support feature analysis and model interpretability, several exploratory visualizations were generated. Figure 10 displays a grouped box plot showing the relationship between Age and DiagnosisType, while Fig. 11 presents a scatter plot of Age versus BMI, color-coded by DiagnosisType and scaled according to PGlucose values. These visualizations helped to confirm class separability and feature relevance, supporting the multiclass classification objective. To further improve classifier performance, each ML model underwent hyperparameter tuning using the PSO technique. PSO played a critical role by iteratively adjusting the positions of particles within the swarm to identify the optimal parameter configuration for each model. This process led to significant improvements in classification accuracy. To enhance model transparency, SHAP summary plots were generated. These plots quantified the contribution of each feature to both individual predictions and the overall model output, identifying the most influential variables and supporting clinical interpretability. In the final stage, the performance of all classifiers was compared to determine the most effective model for recommendation. The classifier achieving the highest predictive accuracy was selected as the optimal model, ensuring maximum generalization to new data. To support this selection, three critical evaluation visualizations were generated: a Confusion Matrix applied to the full dataset to assess class-wise prediction accuracy; ROC Curves for each class, using the One-vs-Rest (OVR) strategy to evaluate discriminatory power; and a Learning Curve illustrating training versus validation accuracy as dataset size increased. These are presented in Figs. 12, 13, 14, 15, 16, 17, 18, 19 and 20 and further detailed in online Appendix C. The selected model using the DTD dataset was used to make predictions on an external diabetes dataset namely diabetes_prediction_dataset include 9 features of 100,000 records. In addition, an external dataset, the diabetes_Dataset, comprising 34 features and 12 distinct diabetes types, is used in place of the previously employed DTD dataset to evaluate the system’s adaptability and generalization across a wider spectrum of diabetes classifications in online Appendix D. All stages of the pipeline were applied without modification, demonstrating the system’s flexibility and robustness when applied to more complex, real-world datasets. Model performance was assessed using multiple evaluation metrics, including Precision, Accuracy, MSE, R2 Score, and AUC. Among all classifiers, ANN demonstrated superior performance, achieving an overall classification accuracy of 99.98%. These models recorded minimal or no misclassifications across all classes and achieved the highest scores across all evaluation metrics.

Fig. 11
Fig. 11
Full size image

Age and BMI group with diagnosis type and PGlucose.

Fig. 12
Fig. 12
Full size image

RF Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 13
Fig. 13
Full size image

KNN Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 14
Fig. 14
Full size image

DT Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 15
Fig. 15
Full size image

GB Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 16
Fig. 16
Full size image

NB Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 17
Fig. 17
Full size image

AB Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 18
Fig. 18
Full size image

LR Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 19
Fig. 19
Full size image

ANN Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

Fig. 20
Fig. 20
Full size image

SVM Confusion Matrix, Multiclass ROC Curve, and Learning Curve.

The Eqs. (7–11) are used to measure the performance evaluation of the classification techniques as Precision Score, Recall Score and F1-Score parameters are calculated28. TN = True Negative, FP = False Positive, FN = False Negative, TP = True Positive. The precision score provides the accuracy of positive diabetes predictions using Eq. (7).

$${\text{Precision }} = \frac{{{\text{TP}}}}{{{\text{TP }} + {\text{ FP}}}}$$
(7)

Recall score is the ratio of correctly predicted positive cases to all using Eq. (8).

$${\text{Recall }} = \frac{{{\text{TP}}}}{{{\text{TP }} + {\text{ FN}}}}$$
(8)

F1-score is the weighted average of Precision and Recall using Eq. (9).

$${\text{F1 }} = \frac{{2{ * }\left( {\text{Precision * Recall}} \right)}}{{{\text{Precision }} + {\text{ Recall}}}}$$
(9)

Accuracy indicates ML classifiers correctness in the diagnosis of whether a patient is diabetic or non-diabetic using Eq. (10).

$${\text{Accuracy }} = \frac{{{\text{TP }} + {\text{ TN}}}}{{{\text{TP }} + {\text{ TN }} + {\text{ FN }} + {\text{ FP}}}}$$
(10)

MSE is a measure of the loss function using Eq. (11).

$${\text{MSE}} = \frac{1}{{{\text{m *}}\mathop \sum \nolimits_{j = 1}^{m} \left( {X_{j} - \hat{X}_{j} } \right)}}$$
(11)

Table 5 represents the efficiency comparisons of multi-class classification algorithms is evaluated based on five factors, including Training Score, MSE, R2 score, Precision, and AUC. The accuracy comparisons achieved by the nine classifiers is illustrated in Fig. 21.

Table 5 A comparison of ML techniques of training score, MSE, R2 score, precision, and AUC.
Full size table
Fig. 21
Fig. 21
Full size image

ML algorithms and accuracy.

The ML techniques are implemented using Jupiter Notebook and the Python programming language. All previous studies have been performed on the PID dataset only for older ages of female’s gender. We need to extend the study to take into consideration the small ages of both genders, male/female, and add other diabetes dataset types that affect multiclassification to make a strong study. The new DTD dataset includes male/female patients with small and old ages to study effectively the diabetes types. In addition, the new input attributes are generated to support the decision-making of the multiclassification process. The aim of creating new features in the DTD dataset is essential to increase the accuracy and high predictive power of a model.

Table 6, Table 7 and Fig. 22 can conclude the supported model outperforms with high accuracy than other algorithms.

Table 6 A comparison of the ML models reviewed and the proposed model.
Full size table
Table 7 A notations of Table 6 and their descriptions.
Full size table
Fig. 22
Fig. 22
Full size image

A comparison between the proposed model and other literature.

Conclusions and future work

Early detection of diabetes poses a significant challenge within healthcare systems, particularly due to the complex nature of the disease and its varying manifestations. This study demonstrates the powerful potential of supervised ML techniques in accurately detecting and classifying multiple types of diabetes. By leveraging a multiclass classification framework trained on the integrated DTD dataset, which combines patient records from Pediatrics, PID, Pone, and Gestational Diabetes sources, the developed ML system effectively classifies Type 1, Type 2, Gestational, and Normal (non-diabetic) cases. A robust preprocessing pipeline, including MICE-based imputation and SMOTE for class balancing, ensured data quality and fairness. Multiple ML algorithms, such as LR, NB, DT, RF, AB, GB, SVM, KNN, and ANN, were evaluated, with model tuning performed using PSO to optimize hyperparameters. The system’s transparency was enhanced through SHAP analysis and detailed performance visualizations (Confusion Matrix, ROC Curves, and Learning Curves), offering clinical interpretability and confirming the model’s reliability. Notably, classifiers like ANN achieved superior performance, recording up to 99.98% accuracy with minimal misclassification. These models outperformed other techniques reported in the literature, including in external validation tasks using two diverse diabetes datasets: one with 100,000 records and another containing 12 distinct diabetes types across 34 features. The system’s ability to generalize effectively across both simple and complex datasets underscores its adaptability, robustness, and real-world clinical applicability. The integration of this ML system into clinical workflows and self-management protocols has the potential to revolutionize diabetes care, empowering healthcare professionals and patients alike with timely, data-driven insights. By accurately identifying indicators such as insulin levels for Type 1 and Type 2 diabetes and POGTT results for Gestational diabetes, the system facilitates early diagnosis and supports proactive disease management. The comprehensive evaluation using multiple performance metrics, including Precision, Accuracy, MSE, R2 Score, and AUC confirms the adequacy and reliability of the developed system, establishing it as a powerful tool for automated, multi-type diabetes diagnosis. Future extensions of this work could include automating diabetes prediction, developing an Android application for diabetes monitoring, integrating genetic algorithms with prediction mechanisms, and exploring enhancements by incorporating medical scan images such as those related to eye and skin diseases. The designed system, equipped with ML multiclassification algorithms, holds promise for predicting and diagnosing various diseases beyond diabetes.